CN100487722C - Method for determining connection sequence of cascade classifiers with different features and specific threshold - Google Patents

Abstract

The present invention proposes a method for determining the connection order and feature thresholds of a cascade of classifiers for a set of different features. The set of cascaded classifiers is used to extract the connected components to be selected from the candidate connected components decomposed from the image. The method includes the following steps: firstly obtain a plurality of connected components by decomposing at least one image as the current sample, and then send the current sample in parallel to the current cascade classifier of each feature for cyclic training, so as to determine each different The connection order of the cascade classifiers of the features and the feature threshold interval. The present invention also proposes a method for obtaining the image to be selected from the image, using the classifier group cascaded according to the aforementioned method, the non-selected connected components can be quickly removed, and more time is spent on calculating possible It is the selected connected component, which not only improves the image processing speed, but also improves the accuracy of image acquisition.

Description

A Method for Determining the Connection Order and Feature Threshold of Cascade Classifiers of Features

technical field

The invention clearly relates to the field of digital image processing, and in particular to a method for determining the connection sequence and feature threshold of a set of cascade classifiers with different features, and using the cascade classifier group formed by the method to obtain selected The method of specifying the image.

Background technique

Text detection and segmentation in natural scenes has many applications. With the increase of high-performance, low-cost, and portable digital imaging equipment, the application of scene text recognition is expanding rapidly. By using a camera connected to a mobile phone, PDA or other dedicated digital devices, we can easily capture the text around us: such as road names, advertisements, traffic warnings, restaurant menus, and so on. Automatic recognition, translation and pronunciation of these texts can be of great help to overseas tourists, visually impaired people and video retrieval programs, meeting handling, etc.

Fully automatic text extraction from images, especially images of natural scenes, remains a challenging problem. The difficulties include: character fonts, sizes, complex backgrounds, non-uniform lighting, shadows, and image noise. Moreover, the requirements for image processing speed are getting higher and higher.

In recent years, the work on text acquisition in natural scene images has developed rapidly. There are currently two classes of methods for extracting text from natural scene images.

The first category is texture-based methods. In "Support Vector Machine-Based Text Detection in Digital Video" published by Shin et al. in 2000, star-shaped pixel templates were used to reveal the intrinsic characteristics of text. In "Getting Text Regions Using Localized Measurement Methods" published in September 2000, P.Clark et al. carefully proposed five localized measurement methods, and combined these measurements to obtain candidate text regions. Frequency-domain methods have also been used to detect text-like textures, for example: Fourier transform of short scan lines, discrete cosine transform, Gabor transform, wavelet decomposition, multi-resolution edge detection. We found that these methods work well for relatively small characters, such as lines of text in menus or documents, because small text usually has a strong texture response. However, for large characters, such as roadsides or shop names, strong texture responses to complex backgrounds can mislead these algorithms, leaving many large characters undetected.

The second type of method is a method based on Connected-Component (CC). Color quantization, mathematical morphological operations, and symmetric neighborhood filtering are often used to decompose the original image into candidate connected components. These algorithms handle both large and small characters efficiently. But how to extract text connected components from candidate connected components, people often use heuristic methods, such as: aspect ratio, alignment and merge analysis, layout analysis, multi-layer connected component analysis. The disadvantage of this type of method is that all heuristic rules are in a fixed order, and the threshold value is a manual input experience value, which is usually unstable and cannot guarantee robust detection results. In addition, a strong classifier (such as Support Vector Machine, SupportVector Machine, SVM) can be used to extract text connected components from candidate connected components. The disadvantage of this method is that all features must be calculated for each connected component, and the amount of calculation And it takes too much time.

Inspired by face detection technology, the present invention uses cascaded classifier groups to extract connected components (for example, text connected components) to be selected from candidate connected components, and can obtain a high detection rate while improving image processing speed .

Contents of the invention

One of the objects of the invention is to propose a connection of cascaded classifiers (h ₁ , h ₂ , ..., h _M ) that determine a set of different features (F ₁ , F ₂ , ..., F _M ) Methods for sequential and feature thresholding. The cascaded classifier group is used to extract the connected components to be selected from the candidate connected components obtained by image decomposition, where different features are related to the images to be selected. The method includes the following steps:

a. Obtain multiple connected components by decomposing at least one image as the current sample, and use the cascade classifier of M different features as the current cascade classifier of each feature, the current sample includes the set of positive examples P And negative example set N, the positive example is the connected component to be selected, and the negative example is the non-selected connected component;

b. Send the current sample in parallel to the current cascade classifier of each feature, and perform one training in the i cycle training, where i is a positive integer of 0<i≤M, and select all the current samples participating in each training in turn The feature corresponding to the maximum false alarm rate in the feature, thereby determining the connection sequence of the cascade classifier of each different feature, wherein the false alarm rate is the connected component that is mistaken for the selected connected component by the cascade classifier in each training The ratio of the number of connected components that are actually non-selected to the number of current negative examples;

c. After each feature is selected, the current sample is sent to the cascade classifier corresponding to the selected feature for training. During this training process, the false alarm rate and detection rate are constantly changing, and according to the feature The allowed minimum detection rate obtains the threshold interval of the feature, thereby determining the feature threshold interval of the cascade classifier of each different feature; the detection rate is the number of selected connected components correctly detected by a cascade classifier ratio to the number of positive examples; and

d. After performing steps b and c, delete the feature selected in step b and the classifier of the feature to update the current feature and the current classifier of each feature, and keep the set of positive examples in this training unchanged and When the threshold interval of the feature is obtained in step c, the cascaded classifier mistakenly regards the connected component as the selected connected component but is actually not selected as the new negative example set to update the current sample for the next cycle training.

Decomposing the image into connected components in the above step a further includes the following steps:

a1. processing said image with a non-linear Niblack thresholding method; and

a2. Decomposing the processed image into connected components.

Among them, the nonlinear Niblack thresholding method adds a statistical order filter to the background filter and the foreground filter of the standard Niblack method.

Another object of the present invention is to provide a method for obtaining images to be selected from images, comprising the following steps:

A. Decompose the image into connected components;

B. Send this connected component to the first stage of a cascade classifier of a group of different features cascaded according to the aforementioned method, which feature is related to the image to be selected, and each cascade classifier discards non-selected Connected components, and output the connected components to be selected to the next classifier; and

C. Combining the connected components to be selected output by the last classifier in the cascade classifier group to form the image to be selected.

Another object of the present invention is to provide a device for acquiring images to be selected from images, the device comprising:

decomposition means for decomposing the image into connected components;

A group of cascade classifiers with different characteristics cascaded according to the aforementioned method, the connected components are input into the first stage of the cascade classifier, each cascade classifier discards non-selected connected components, and forwards to the next The level classifier outputs the connected components to be selected; and

The image synthesis device is used for combining the connected components to be selected output by the last classifier in the cascade classifier group to form the image to be selected.

Because the method of the invention uses a new nonlinear Niblack method to process the original image, the grayscale image can be efficiently decomposed into multiple candidate connected components, and the quality of the connected components is improved. In addition, the cascaded classifier group trained by the above method can easily remove most non-text connected components, and quickly focus on the connected components that may be considered as text. In this way, the calculation amount of the method is reduced, the image processing speed is improved, and a high detection rate can be obtained.

Description of drawings

Fig. 1 is a flow chart of a method for determining a connection sequence and a feature threshold of a cascade classifier of a set of different features according to an embodiment of the present invention;

FIG. 2 is a flow chart of a method for obtaining a text image from an image according to a second embodiment of the present invention; and

Fig. 3 is a diagram of an apparatus for acquiring text images from images according to a third embodiment of the present invention.

Detailed ways

The present invention will be further described in detail in conjunction with the accompanying drawings and specific embodiments.

As mentioned above, the method of the present invention is inspired by face detection technology, and uses a cascade classifier group to extract text connected components from candidate connected components. The candidate connected components are obtained by decomposing the original image, which can be a natural scene image . The text connected components are combined to form a text image, so that we can obtain text images from natural scene images.

So, how can the aforementioned group of cascaded classifiers extract textual connected components from candidate connected components?

First, we propose 12 different features, which can effectively distinguish textual or non-textual connected components. These 12 features are then corresponding to each classifier in the cascade classifier group, and the cascade classifier group is trained to determine the connection sequence and feature threshold of the cascade classifiers of this group of different features. Such a cascaded cascade classifier group can quickly discard non-text connected components and output text connected components.

Next, we describe in detail the 12 features that reveal the intrinsic properties of connected components of text, including: geometric features, edge contrast features, shape regular features, stroke features, and spatial consistency features.

1. Geometric features

Geometric features include Area Ratio, Length Ratio and Aspect Ratio. They are very effective at rejecting large numbers of apparently non-textual connected components with little computational cost. Therefore they can drastically reduce the execution time of the entire algorithm.

The area ratio AreaRatio is the ratio of the area of the connected component to the image area, which is used to exclude connected components that are too large or too small. The formula is:

$\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; AreaRatio</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; picture</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

The length ratio LengthRatio is used to exclude connected components that are too long or too short:

$\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; LengthRatio</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; max</span>}<span\; class="notranslate">\; \{</span>\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; \}</span>}{\mathrm{<span\; class="notranslate">\; max</span>}<span\; class="notranslate">\; \{</span>\mathrm{<span\; class="notranslate">\; Pic</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; PIC</span>}<span\; class="notranslate">\; \}</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

The aspect ratio AspectRatio is used to propose too thin connected components:

$\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; AspectRatio</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; max</span>}<span\; class="notranslate">\; \{</span>\frac{\mathrm{<span\; class="notranslate">\; width</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; height</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; ,</span>\frac{\mathrm{<span\; class="notranslate">\; height</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; width</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; \}</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

In (2) and (3) above, w represents the bounding box width of the connected component, h represents the height of the bounding box of the connected component, W represents the width of the image, and H represents the height of the image.

2. Edge contrast feature

The edge contrast feature includes edge contrast (Edge Contrast), which is the ratio of the coincidence degree of the boundary of the connected component and the edge image of the original image to the boundary of the connected component, and its formula is:

$\mathrm{<span\; class="notranslate">\; EdgeContrast</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; Border</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&cap;</span>\mathrm{<span\; class="notranslate">\; Edge</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; picture</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; Border</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

Among them, Border (CC) is the boundary pixel of connected components, Edge (Picture) is the edge detection image of the original image, which is the union of Canny operator and Sobel operator, and its formula is:

Edge(Picture)＝Canny(Picture)∪Sobel(Picture) (5)

The edge contrast feature is the most important feature. This feature is proposed based on a very general perspective, regardless of complex backgrounds and non-uniform lighting, text connected components are usually "height surrounded" by their edge responses. Therefore, we use Equation (4) to measure the degree of edge envelopment of a connected component. This feature makes full use of the advantages of texture-based detection algorithms, and it also has a strong response to large characters. Moreover, this feature provides an image-independent measure of the edge contrast of each connected component.

3. Shape regularization features

Text connected components tend to have more regularized shapes than noisy connected components in natural scenes. Based on this viewpoint, we propose 4 features: number of voids, profile roughness, compactness, and duty cycle. We find that text connected components have small values for hole number and contour roughness, and large values for compactness and duty cycle; the opposite is true for non-text connected components. These features are used to suppress connected components with irregular shapes but strong texture responses.

$\mathrm{<span\; class="notranslate">\; Featuer</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; Contour\; Roughness</span>}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; open</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; imfill</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; |</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; )</span>$

Feature_CCHoles＝|imholes(CC)| (7)

$\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; Compact</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span>}{{<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; Perimeter</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 8</span>}<span\; class="notranslate">\; )</span>$

$\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; OccupyRatio</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; BoundingBox</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 9</span>}<span\; class="notranslate">\; )</span>$

In the above formula, imfill (CC) is the operation of filling the internal holes of the connected components, 2×2 is the structure element of the morphological opening operation, and the morphological opening operation (open) is a smoothing operation on the connected components.

4. Statistical characteristics of strokes

Characters are composed of strokes, so we propose two computationally complex features to reveal the stroke statistics of connected components. These two features actually check the "irregularity" of connected components in terms of character strokes.

The first feature is the average stroke width MeanStrokeWidth. We are based on the idea that the stroke width of characters is usually relatively small:

Feature_Stroke_Mean＝Mean(strokeWidth(skeleton(CC))) (10)

The second feature is the normalized stroke width standard deviation, we are based on the idea that strokes of the same character tend to have similar widths, and connected components with very large values on the stroke width standard deviation feature are more likely to be noise:

$\mathrm{<span\; class="notranslate">\; Featuer</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; Stroke</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; Std</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; Deviation</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; strokeWidth</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; skeleton</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; mean</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; strokeWidth</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; skeleton</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 11</span>}<span\; class="notranslate">\; )</span>$

In the above formula, skeleton is a morphological skeleton algorithm, which extracts the skeleton of connected components to obtain a skeleton diagram, strokeWidth is the stroke width obtained for each point on the skeleton diagram, and Mean is the average value of all points on the skeleton diagram , to get the average width.

5. Spatial Consistency Features

The last two spatial consistency features explore spatial consistency information to filter out non-text connected components. Noise tends to have less spatial regularity and aggregation, so we propose these two features. Spatial coherence features include spatial coherence area ratio (Spatial Coherence Area Ratio) and spatial coherence boundary features (Spatial CoherenceBoundary Touching), where,

$\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; AreaRatio</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; imdilate</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; picture</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 12</span>}<span\; class="notranslate">\; )</span>$

Feature_Boundary_S＝Bound(imdilate(CC, 5×5)) (13)

In the above formula, imdilate is the operation of expanding the connected components, and 5×5 is the structure element of the expansion operation.

In the case where many obviously non-text connected components have been excluded, in each layer (Niblack has two color layers of black and white), if after a small structural element is expanded, a certain connected component expands very strongly , then it is likely to be spatially correlated random noise. This is not the case for text connected components. Due to the structural nature of strings, there is often a little space between characters. After expansion, they will not stick to each other and expand into a large connected component. By using a spatially consistent filter, we can effectively reduce image noise.

After the above 12 features that can effectively distinguish text or non-text connected components are proposed, the 12 features are corresponding to each classifier in the cascade classifier group, and the cascade classifier group is trained. Our training method needs to solve two problems, first, in what order to arrange these features; second, what should be the threshold value on each feature. The advantage is that the cascaded classifier group can be cascaded in a weak first and then strong manner, which not only ensures the accuracy of image acquisition, but also improves the image processing speed. FIG. 1 is a flow chart of a method for determining the connection sequence and feature threshold of a cascade classifier of a set of different features according to the first embodiment of the present invention.

To perform training, firstly a training example is determined (step 110). For example, we can randomly select 200 pictures from the picture library, and decompose these 200 pictures into multiple connected components as training samples (the method of decomposing the original picture into connected components will be described in detail below). The training sample rate includes positive example set P and negative example set N. Positive examples are connected components that we manually label as text, and negative examples are connected components that we manually label as non-text.

For each training sample (that is, a connected component), it has two Boolean values: one is the groundtruth value (GroundTruth), that is, whether this sample is text, true is text, false is non-text; the other is classification The output value of the classifier, that is, whether the classifier considers this sample to be text, the output positive is text, and the negative is non-text. In this sense, the false positive rate indicates the ratio of non-text samples that are mistaken for text by the classifier to all non-text samples; the detection rate is actually true-positive, which means that the classifier correctly considers it to be text The ratio of text samples to all text samples; false-rejection is false-negative, indicating the ratio of non-text samples that are correctly considered not to be text by the classifier to all non-text samples.

As a set of positive examples, P does not change during the entire training process, because we expect each positive example (text connected component) to pass through all classifiers, that is to say, each classifier must "know" these positive examples, that is, to Learn them. For the set of negative examples N, since each classifier will "intercept" a part of the negative examples, for each classifier in the cascade, the negative examples they see are different. The first classifier sees all negative examples, the second only sees those negative examples that were misclassified as text by the first one... From the perspective of the latter classifier, it only needs to pay attention to the previous classifiers There is no problem that can be correctly distinguished, which means that the counterexamples it has to deal with are only non-text connected components that pass through all previous classifiers. So, we need to change the set of negative examples N in each iteration of training.

As mentioned above, the method of decomposing the original picture into connected components is described in detail next.

As we all know, decomposing an image into connected components is a very critical step in connected component-based methods. If the decomposition step achieves poor results, the performance of the entire algorithm will drop dramatically. Existing decomposition methods mainly pursue efficiency and robustness.

This embodiment uses a new decomposition method for decomposing an image into connected components, which includes two steps: first, the original image is processed with a nonlinear Niblack thresholding method; and then the processed image is decomposed into connected components.

The key to the Niblack method is that it believes that the intensity of those text pixels that people care about will have a certain gap with the average intensity of its neighborhood, and this gap is greater than k times the standard deviation of the intensity of its neighborhood. It was originally used to binarize images. In this embodiment, we use this method to process the image first, and then decompose the processed image into candidate connected components, which can achieve high efficiency and implementation on the basis of existing effectiveness and robustness. low complexity.

Among them, the nonlinear Niblack thresholding method adds a statistical order filter to the background filter and the foreground filter of the standard Niblack method. The formula for the nonlinear Niblack thresholding method is:

$\mathrm{<span\; class="notranslate">\; NLNiblack</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{c}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; <</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{<span\; class="notranslate">\; +</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>\\ <span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; <</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>\\ \begin{array}{cc}\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; other</span>}\end{array}\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 14</span>}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; T</span>}}_{<span\; class="notranslate">\; \&PlusMinus;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; B</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&PlusMinus;</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; \&Center\; Dot;</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; )</span>$

${\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; order</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; mean</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; B</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; B</span>}}<span\; class="notranslate">\; ]</span>$

${\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; order</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; Deviation</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; ]</span>$

Where: k is an empirical value according to the standard Niblack method, and is set to a value between 0.17-0.19, preferably, it is set to 0.18 in this embodiment. f(x, y) is the pixel intensity at the (x, y) position of the input image, Mean(, W) is the mean filter with a window width of W, and Deviate(, W) is the standard deviation with a window width of W Filter, Order[, p, W] is an order statistical filter with p as the percentage and W as the width.

中，滤波器宽度W_B设为原始图像宽度的1/16，百分比p1设为50％。这是因为大的中值滤波器可以在提取背景对象的同时不排除它们的高频分量。这个背景滤波器可以应付自然场景中的非均匀光照情况。In this embodiment, in the background filter

, the filter width W _B is set to 1/16 of the original image width, and the percentage p1 is set to 50%. This is because a large median filter can extract background objects without excluding their high-frequency components. This background filter can cope with non-uniform lighting in natural scenes.

中，宽度W_F是W_B的1/5，p2设为80％。对于具有较大方差的小块区域，这个高百分比的滤波器可以有效地将其影响传播到邻近的区域，同时能有效地抑制局部噪声。in the foreground filter

, the width W _F is 1/5 of W _B , and p2 is set to 80%. For small patch regions with large variance, this high percentage of filters can effectively propagate its influence to neighboring regions while effectively suppressing local noise.

Of course, the above filter width and percentage can be adjusted according to actual needs.

In addition, it is worth mentioning that the above-mentioned image decomposition step can also not use the nonlinear Niblack method to process the image, but use the existing technology of decomposing the image into connected components, which can also achieve the purpose of the present invention, but because the existing The quality of the connected components obtained by the technology is poor, so the overall effect of this method will also be reduced.

Next, setup and initialization operations are performed (step 120).

Set the overall system target detection rate D _target of the cascaded classifier group (h ₁ , h ₂ , ... h ₁₂ ) = 0.95; and manually input the target detection rate.

Initialize variables: set the overall detection rate D ₀ =1.0, the set of negative examples N ₁ =N, the number of cycles i=0, the range of i is 0<i≤M, that is, 0<i≤12, and initialize the feature set, the feature set Contains 12 features (F ₁ , F ₂ , . . . F ₁₂ ). There is a one-to-one correspondence between classifiers and features.

Let the number of loops i=i+1 (step 130).

It is judged whether i is greater than M (step 140). If i is not greater than M, perform one of i loop calculations. For example, if i=1, then the first cycle calculation is performed. The following takes the first cycle calculation as an example to describe in detail.

The samples in the positive example set P and the current negative example set _N1 are sent to each cascaded classifier in parallel for training (step 150). Each classifier computes feature values for all examples. For example, if the feature corresponding to the first classifier is the geometric feature "area ratio", then the area ratio of all samples is calculated, that is, the ratio of the area of the bounding box of the connected component of the sample to the area of the entire image.

After obtaining the eigenvalues of all samples, take the eigenvalues as the abscissa and the number of connected components as the ordinate to form the eigenvalue distribution diagram of the positive example P and the negative example N ₁ .

For each feature, set a threshold interval with an initial value of (-∞, +∞). If the feature value of a sample is outside the threshold interval, the sample will be judged as Connected components of non-text; if the feature value of a sample is outside the threshold interval, the sample is judged as a connected component of text by the cascade classifier corresponding to the feature.

In this threshold interval (-∞, +∞), all samples conform to this threshold interval, therefore, the detection rate d of each classifier is 1, and the false alarm rate f is also 1. For each feature, the threshold interval is continuously reduced, so that the feature values of more and more samples do not meet the threshold interval, positive examples and negative examples are continuously judged as non-text connected components, and the detection rate of each cascade classifier d _1j and the false alarm rate f _1j continue to decrease, when the detection rate d _1i of a certain classifier in the first cycle training drops to not less than the overall detection rate D _i-1 after the previous cycle, stop narrowing the threshold interval . Here D _i-1 =D ₀ =1.0. Due to the discrete nature of the distribution during actual calculation, it is impossible for d _1i to be reduced to be equal to D ₀ , only slightly larger.

At this threshold interval, the detection rate d _1j , the false alarm rate f _1j and the probability FR _1j of correctly discarding non-text connected components of each cascaded classifier are calculated, where FR _1j =1-f _1j is a level The ratio of the number of non-text connected components correctly discarded by the joint classifier to the number of current negative examples.

In the current feature set, that is, among the 12 features, select the feature feature _k corresponding to the maximum false alarm rate f _1j (step 160). The selected feature feature _k is the first feature, and its corresponding classifier is the first classifier of the cascaded classifier group.

The feature corresponding to the maximum false alarm rate is selected because it can be seen from the above-mentioned round of calculations that under the same conditions, the feature corresponding to the maximum false alarm rate most considers non-text samples as text samples, then the feature It is considered to be the least effective feature, and its classification ability is the worst, so it should be placed at the top of the cascaded classifier group, and so on, so that the classifier group cascaded by this method has the weakest Strong cascading approach.

Next, calculate the quality ratio of the selected feature feature _k and its allowed minimum detection rate (step 170).

The quality ratio of the selected feature feature _k γ=FR _k /∑FR _1j , where FR _k is the probability that the cascaded classifier corresponding to the selected feature feature _k in the first round of training correctly discards non-text connected components, Equivalent to the quality of the classifier, this value is obtained in step 160; ∑FR _1j represents the sum of the probability that the cascaded classifiers corresponding to all features in the first round of training correctly discard non-text connected components. The ratio between the two is the quality ratio of the cascade classifier corresponding to the selected feature, which is used to measure the ability of the feature to distinguish text connected components from non-text connected components.

According to the detection rate distribution formula d _i =(D _target /D _i-1 ) ^γ , calculate the allowed minimum detection rate d _min of feature _k , where D _i-1 is the overall detection rate after the previous cycle training, and i is Cycles. Since it is the first cycle training, here D _i-1 =D ₀ =1.0, d _min =(D _target ) ^γ .

How the detection rate distribution formula is obtained is described in detail below.

Assume that we will serially send some connected components into a set of cascade classifiers with M different features, and classify them level by level. If any classifier thinks that a connected component is a non-text connected component, it will be removed. If It is considered as a text connected component, that is, it is output to the next classifier for classification again. In this way, we can easily obtain the following relationship:

$f=\underset{i=1}{\overset{m}{\mathrm{\&Pi;}}}{f}_i$ $\mathrm{D.}=\underset{i=1}{\overset{m}{\mathrm{\&Pi;}}}{d}_i---\left(15\right)$

则对应的虚警率为

总体检测率 $D\&prime;=\underset{i=1}{\overset{M}{\mathrm{\&Pi;}}}{d}_i\&prime;,$ 总体虚警率为 $F\&prime;=\underset{i=1}{\overset{M}{\mathrm{\&Pi;}}}{f\&prime;}_i.$ 在D＝D′的情况下，未必有F＝F′。我们的目的是，在总体检测率D＝D_target的情况中，选择具有最小虚警率F的那组检测率向量。那么如何在D固定的情况下，最小化F呢？For each of the M classifiers, there is a detection rate d _i , and for this d _i there is a false alarm rate f _i , in order to simplify the expression, we form d _i into a vector {d ₁ , d ₂ , ..d _M }, the overall detection rate at this time $\mathrm{D.}=\underset{i=1}{\overset{m}{\mathrm{\&Pi;}}}{d}_i,$ The overall false alarm rate $f=\underset{i=1}{\overset{m}{\mathrm{\&Pi;}}}{f}_i.$ If for these M classifiers we set another set of detection rates

Then the corresponding false alarm rate is

overall detection rate $\mathrm{D.}\&prime;=\underset{i=1}{\overset{m}{\mathrm{\&Pi;}}}{d}_i\&prime;,$ The overall false alarm rate $f\&prime;=\underset{i=1}{\overset{m}{\mathrm{\&Pi;}}}{f\&prime;}_i.$ In the case of D=D', it is not necessary that F=F'. Our purpose is to select the set of detection rate vectors with the smallest false alarm rate F in the case of the overall detection rate D=D _target . So how to minimize F when D is fixed?

By log-transforming the basic form of Equation (15), we find that the overall detection rate is distributed linearly across all classifiers:

$\log \left(f\right)=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\log \left(f_i\right)$ $\log \left(\mathrm{D.}\right)=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\log \left(d_i\right)---\left(16\right)$

Assuming that the overall detection rate D is linearly distributed to all classifiers according to the "quality" of the classifier, the "quality" of the i-th classifier is Q, and the sum of the quality of all classifiers is $Q=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{Q}_i,$ The quality ratio _γi of the i-th classifier is defined as:

${\mathrm{<span\; class="notranslate">\; \&gamma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 17</span>}<span\; class="notranslate">\; )</span>$

Let D be the overall detection rate, we can express the distribution formula as follows, the detection rate d _i assigned to the i-th classifier is:

${\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; )</span>}^{{\mathrm{<span\; class="notranslate">\; \&gamma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 18</span>}<span\; class="notranslate">\; )</span>$

From equation 1) we have:

$\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Pi;</span>}}}{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Pi;</span>}}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; )</span>}^{{\mathrm{<span\; class="notranslate">\; \&gamma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; D.</span>}}^{\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; \&gamma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; D.</span>}}^{\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\frac{{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{\mathrm{<span\; class="notranslate">\; Q</span>}}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 19</span>}<span\; class="notranslate">\; )</span>$

This shows that my allocation algorithm is numerically correct in the first place.

Because D∈[0,1], its exponential function is a monotonically decreasing function. The better the "quality" of a classifier and the larger γ, the smaller the assigned detection rate d. Because "quality" means that this classifier can most effectively exclude non-text, so we allow its detection rate d to be smaller, so that it can have more space to exclude non-text connected components. Decreasing the detection rate means setting more stringent conditions, so that more non-text connected components can be excluded. The "quality" of a classifier can be measured by the probability of correctly excluding non-text.

After obtaining the minimum detection rate allowed by the selected feature featurek, send all samples in the positive example set P and the current negative example set _N1 to the cascaded classifier h _k corresponding to the selected feature for training (step 180) .

This classifier computes feature values for all examples. For example, if the feature is a length ratio, the length ratios of all samples are calculated, and the calculation formula refers to the above description.

Set a threshold interval with an initial value of (-∞, +∞). When the feature value of a sample is outside the threshold interval, the sample is judged as a non-text connected component by the cascade classifier h _k .

The threshold range is continuously narrowed, so that positive examples and negative examples are continuously judged as non-text connected components, and the detection rate d _k and false alarm rate f _k of the cascade classifier h _k continue to decrease. When d _k drops to no less than step 180 When the allowed minimum detection rate d _min obtained in , stop narrowing the threshold interval; the threshold interval at this time is the threshold interval of the selected feature _k .

So far, the work of selecting features and determining feature threshold intervals has been completed.

Next, the variables are to be updated for the next cycle training (step 190).

Delete the selected feature and the classifier of the feature to update the current feature set and the current classifier of each feature. The non-text connected component that was mistaken for the text connected component by the cascade classifier when the threshold interval of the feature was obtained in step 180 is used as a new negative example set N _i+1 , and the positive example set P remains unchanged, thereby updating the current example. Then update the current overall detection rate D _i =D _i-1 *d _min for the next cycle training.

The next cycle calculation is exactly the same as the first one above, each time a feature is selected and the threshold interval of the feature is obtained. The serial number of the classifier corresponding to each selected feature is the number i of the cycle. Until i is greater than M, the loop calculation ends.

The cascaded classifier group cascaded according to the connection order determined by the above method can quickly exclude non-text connected components, and spend more time on computing connected components that may be text.

The features proposed in this embodiment are related to text images, and can effectively distinguish text or non-text connected components. Therefore, the cascaded classifier group corresponding to this set of features can obtain text connected components from candidate connected components, thus, by combining Text connected components to obtain the text image we need. However, those skilled in the art should know that if the proposed features are related to other content to be selected, which can be any content we want to obtain from the original image, then the set of cascaded classifiers corresponding to this set of features The connected components to be selected can be obtained from the candidate connected components, so as to form the image we want to select, not limited to the text image. Therefore, the cascaded classifier group determined by the method in this embodiment can obtain the connected components to be selected according to the features related to the content to be selected.

Fig. 2 is a flow chart of a method for acquiring a text image from an image according to a second embodiment of the present invention.

First, the original image is decomposed into multiple candidate connected components (step 210). The original image here can be a natural scene image. In this step, the original image can be processed with a nonlinear Niblack thresholding method; and then the processed image can be decomposed into multiple connected components. Here, the method of processing the original image with the nonlinear Niblack thresholding method is the same as that in the first embodiment, and will not be repeated here. Candidate connected components can be obtained quickly and robustly with nonlinear Niblack thresholding method.

Secondly, a plurality of candidate connected components are sent to the first stage of a cascade classifier of a group of different features cascaded according to the method of the first embodiment, the feature is related to the text image, and each cascade classifier discards non-text connected components, and output text connected components to the next classifier (step 220). The connection sequence and feature threshold of the cascade classifier group are determined according to the method of the first embodiment.

Specifically, after multiple candidate connected components are input to the first stage of the cascaded classifier group, the first classifier calculates the feature values of all received connected components according to its corresponding features. Compare the eigenvalues of all connected components with the threshold interval of the feature; finally discard the connected components whose eigenvalues are outside the threshold interval as non-text connected components; use the connected components whose eigenvalues are within the threshold interval as text The connected components are output to the second-level classifier. That is to say, the connected components rejected by the first classifier will no longer be input into the second classifier, and further calculation and judgment are not required, therefore, a lot of calculation time can be saved.

After the second classifier receives the connected components output by the first classifier, it performs the same calculation and classification work, and so on, until the last classifier discards the non-text connected components and outputs the text connected components.

Optionally, the text connected components output by the cascaded classifier group can be further input into a strong classifier (step 230). The strong classifier is a classifier trained by the standard Adaboost method, and the features of the strong classifier are the same as those of the aforementioned cascade classifier group. The strong classifier linearly combines all the eigenvalues of each connected component output by the cascaded classifier group and judges whether the connected component is a text connected component, thereby discarding the non-text connected components and outputting the text connected components. Since all the eigenvalues of each connected component have been calculated in the cascade classifier group, the total eigenvalues of the connected component can be obtained only by linear combination in this strong classifier. In this way, less computation time can be spent and the accuracy can be further improved.

Of course, the purpose of the present invention can be achieved without using a strong classifier here, and the accuracy of the final image can be further improved by adding a strong classifier.

Finally, the text connected components output in step 230 are combined to form a text image (step 240). This way, we get the text image we need from the original image.

In the method of this embodiment, since a new nonlinear Niblack method is used to process the original image, the grayscale image can be efficiently decomposed into multiple candidate connected components, and the quality of the connected components is improved. In addition, cascaded classifier groups are able to easily remove most non-text connected components and quickly focus on connected components that are considered likely to be text. In this way, the calculation amount of the method is reduced, the image processing speed is improved, and a high detection rate can be obtained.

Those skilled in the art should be well aware that although the features of the cascaded classifier group in this embodiment are related to text images, this feature can also be related to other selected content, so the method in this embodiment can also be used for Get any image to select from images, not limited to text images.

Fig. 3 is a diagram of an apparatus for acquiring text images from images according to a third embodiment of the present invention. The device 300 includes a decomposition device 310 , a cascaded set of classifiers 320 , a strong classifier 330 and an image synthesis device 340 .

Decomposing means 310 is used to decompose the original image into multiple connected components. The decomposing device 310 also includes a processing device 312 and an image decomposing device 316 . The processing device 312 uses the nonlinear Niblack thresholding method to first process the original image, and the nonlinear Niblack thresholding method here is the same as that in the first embodiment. The image decomposition means 316 decomposes the processed image into a plurality of connected components.

The cascaded classifier group 320 is a group of cascaded _classifiers ( _{h 1} , _{h 2 ,} .. .h ₁₂ ), these features are related to text images. This connected component is input to the first stage of the set of cascaded classifiers, and each cascaded classifier discards the non-text connected component and outputs the text connected component to the next classifier.

Computing means, comparing means and output means are also included in each classifier. The calculation device is used to calculate the received feature values of all the connected components according to the features corresponding to the classifier. The comparing means compares the eigenvalues of all the connected components with the threshold interval of the feature respectively. The output device discards the connected components whose feature values are outside the threshold interval as non-text connected components; outputs the connected components whose feature values are within the threshold interval as text connected components to the next-level classifier.

Strong classifier 330, the strong classifier is a classifier trained by the standard Adaboost method, the feature of the strong classifier is the same as the feature of the cascade classifier group 320, that is, the feature of the strong classifier includes the cascade classifier group All the features of the 320. The strong classifier 330 linearly combines all the feature values of the connected components output by the cascade classifier group 320, and judges whether the connected components are text connected components, thereby discarding the non-text connected components and outputting the text connected components.

Image synthesis means 340, configured to combine the text connected components output by the strong classifier 330 to form a desired text image.

Those skilled in the art should be well aware that although the features of the cascaded classifier group 320 in this embodiment are related to text images, this feature can also be related to other selected content, so the device in this embodiment can also be used For getting any image to select from image, not limited to text image.

The present invention has been described in detail in conjunction with the above-mentioned typical embodiments, and various options, modifications, changes, improvements and/or basic equivalent technologies, currently known or (possibly) unknown, are well known to those skilled in the art of. Therefore, the above-mentioned exemplary embodiments of the present invention are intended to illustrate but not limit the present invention. Various changes may be made without departing from the spirit and scope of the invention. Accordingly, the present invention may embrace all known or later developed alternatives, modifications, variations, improvements and/or substantially equivalent techniques.

Claims (44)

1. A method for determining the connection order and feature threshold of a cascade of classifiers (h ₁ , h ₂ , ..., h _M ) for a set of different features (F ₁ , F ₂ , ..., F _M ) , the cascaded classifier group formed by the method is used to obtain the image to be selected from the image, and the different features are related to the image to be selected, wherein M is a positive integer ≥ 1, characterized in that it includes The following steps: a. Obtain multiple connected components by decomposing at least one image as the current sample, and use the cascade classifier of M different features as the current cascade classifier of each feature, the current sample includes the set of positive examples P And a negative example set N, the positive example is marked as a connected component to be selected, and the negative example is marked as a non-selected connected component; b. Send the current sample in parallel to the current cascade classifier of each feature, and perform one training in the i cycle training, where i is a positive integer of 0<i≤M, and select all the current samples participating in each training in turn The feature corresponding to the maximum false alarm rate in the feature, thereby determining the connection sequence of the cascade classifier of each different feature, wherein the false alarm rate is the connected component that is mistaken for the selected connected component by the cascade classifier in each training The ratio of the number of connected components that are actually non-selected to the number of current negative examples; c. After each feature is selected, the current sample is sent to the cascade classifier corresponding to the selected feature for training. During this training process, the false alarm rate and detection rate are constantly changing, and according to the feature The allowed minimum detection rate obtains the threshold interval of the feature, thereby determining the feature threshold interval of the cascade classifier of each different feature; the detection rate is the number of selected connected components correctly detected by a cascade classifier ratio to the number of positive examples; and d. After performing steps b and c, delete the feature selected in step b and the classifier of the feature to update the current feature and the current classifier of each feature, and keep the set of positive examples in this training unchanged and When the threshold interval of the feature is obtained in step c, the cascaded classifier mistakenly regards the connected component as the selected connected component but is actually not selected as the new negative example set to update the current sample for the next cycle training. 2. The method according to claim 1, wherein, decomposing the image into connected components in step a further comprises the steps of: a1. process the image with nonlinear Niblack thresholding method; a2. Decomposing the processed image into connected components. 3. The method according to claim 2, wherein the nonlinear Niblack thresholding method has respectively increased a statistical order filter in the background filter and the foreground filter of the standard Niblack method. 4. the method for claim 3, is characterized in that, the formula of described nonlinear Niblack thresholding method is: $\mathrm{<span\; class="notranslate">\; NLNiblack</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ></span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{<span\; class="notranslate">\; +</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>\\ <span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; <</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>\\ \mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; other</span>}\end{array}\right.$ ${\mathrm{<span\; class="notranslate">\; T</span>}}_{<span\; class="notranslate">\; \&PlusMinus;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; B</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&PlusMinus;</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; \&Center\; Dot;</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; )</span>$ ${\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; order</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; mean</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; B</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; B</span>}}<span\; class="notranslate">\; ]</span>$ ${\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; order</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; Deviation</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; ]</span>$ Where: k is set to 0.17-0.19 according to the standard Niblack method; f(x, y) is the pixel intensity at the (x, y) position of the input image; Mean(, W) is a mean filter with a window width of W; Deviation (, W) is a standard deviation filter with a window width of W; Order[, p, W] is an order statistical filter with p as the percentage and W as the width. 5. The method of claim 4, wherein the background filter , the filter width W _B is set to 1/16 of the image width, and the percentage p1 is set to 50%; in the foreground filter , the width W _F is 1/5 of W _B , and p2 is set to 80%. 6. The method according to claim 5, characterized in that setting and initializing operations are carried out before step b carries out cycle training, further comprising the following steps: Setting the system overall target detection rate D _target of the cascaded classifier group (h ₁ , h ₂ , ... h _j ), where 0<j<=M, M>1; Initialization variables: overall detection rate D ₀ =1.0, set of negative examples N ₁ =N, number of cycles i=0, range of i is 0<i≤M, and initialization feature set, which contains j features (F ₁ , F ₂ , . . . F _j ), 0<j<=M, M>1. 7. the method for claim 6 is characterized in that, the cycle training in the step b further comprises the following steps: b1. Send the samples in the positive example set P and the current negative example set Ni to each cascade classifier in parallel, and calculate the feature values of all samples, b2: For each feature, set a first threshold interval with an initial value of (-∞, +∞). When the feature value of a sample is outside the first threshold interval, the sample is excluded from the first threshold interval. The cascade classifier corresponding to the feature is judged as a non-selected connected component; b3: For each feature, the first threshold interval is continuously reduced, so that positive examples and negative examples are continuously judged as non-selected connected components, and the detection rate d _ij and false alarm rate f _ij of each cascade classifier are continuously Decline, when the detection rate d _i of a certain classifier of the i-th cycle training drops to not less than the overall detection rate D _i-1 after the previous cycle, stop shrinking the first threshold interval; and b4: Obtain the detection rate d _ij of each cascaded classifier, the false alarm rate f _ij and the probability FR _ij of correctly discarding non-selected connected components when obtaining the current first threshold interval, where FR _ij = 1-f _ij , as the ratio of the number of non-selected connected components correctly discarded by a cascade classifier to the current number of negative examples; and b5. In the current feature set, select the feature feature _k with the largest false alarm rate _fij , and the sequence number of the cascade classifier corresponding to the selected feature feature _k is the current cycle number i. 8. The method according to claim 7, further comprising the following steps after step b5: b6. According to the result of step b5, calculate the mass ratio of the selected feature feature _k γ=FR _k /∑FR _ij , where FR _k is the cascade corresponding to the selected feature feature _k in the i-th cycle training The probability that the classifier correctly discards the non-selected connected components is equivalent to the quality of the cascaded classifier, and ∑FR _ij is the sum of the probability of correctly discarding the non-selected connected components of all cascaded classifiers in the ith cycle training; as well as b7. According to the detection rate allocation formula d _i =(D _target /D _i-1 ) ^γ , calculate the allowed minimum detection rate d _min of the feature feature _k , where D _i-1 is the overall detection rate after the previous cycle training , i is the number of cycles; and b8. Update the current overall detection rate D _i =D _i-1 *d _min . 9. The method according to claim 8, characterized in that, the detection rate allocation formula in the step b7 is obtained by the following method: Assuming that the overall detection rate D is linearly distributed to all cascade classifiers according to the "quality" of the cascade classifiers, the quality of the i-th cascade classifier is Q _i , and the sum of the qualities of all cascade classifiers is $Q=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{Q}_i,$ The mass ratio _γi of the i-th cascade classifier is defined as: ${\mathrm{<span\; class="notranslate">\; \&gamma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}}$ Then the detection rate allocation formula is expressed as follows, that is, the detection rate d _i assigned to the i-th classifier is: ${\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; )</span>}^{{\mathrm{<span\; class="notranslate">\; \&gamma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}$ 10. The method of claim 9, wherein step c further comprises the steps of: c1. Send the samples in the positive example set P and the current negative example set Ni to the cascade classifier h _k corresponding to the selected features, and calculate the feature values of all samples; c2. Set an initial value as the second threshold interval of (-∞, +∞), when the feature value of a sample is outside the second threshold interval, the sample is classified by the cascade classifier h _k is judged as a non-selected connected component; and c3. The second threshold interval is continuously reduced, so that positive examples and negative examples are continuously judged as non-selected connected components, and the detection rate d _k and false alarm rate f _k of the cascade classifier h _k continue to decrease, when When d _k drops to not less than the allowed minimum detection rate d _min obtained in step b7, stop narrowing the second threshold interval; the second threshold interval at this time is the threshold interval of the selected feature _k . 11. The method according to any one of claims 1-10, wherein the image to be selected is a text image, the connected components to be selected are text connected components, and the non-selected Connected components are non-text connected components, and features related to text images include: geometric features, edge contrast features, shape regular features, stroke features, and spatial consistency features. 12. The method of claim 11, wherein the geometric features include area ratios, length ratios and aspect ratios, wherein, The area ratio is the ratio of the area of the connected components to the image area, and its formula is: $\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; AreaRatio</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; picture</span>}<span\; class="notranslate">\; )</span>}$ The formula for the length ratio is: $\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; MLengthRatio</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; max</span>}<span\; class="notranslate">\; \{</span>\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; \}</span>}{\mathrm{<span\; class="notranslate">\; max</span>}<span\; class="notranslate">\; \{</span>\mathrm{<span\; class="notranslate">\; Pic</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; PIC</span>}<span\; class="notranslate">\; \}</span>}$ The formula for aspect ratio is: Feature_AspectRatio=max{w/h, h/w} In the above formula, CC represents the connected component, max{a, b} represents the larger value of a and b, w represents the width value of the bounding box of the connected component, h represents the height value of the bounding box of the connected component, PicW Indicates the width of the image, and PicH indicates the height of the image. 13. The method according to claim 11, wherein the edge contrast feature comprises edge contrast, and the edge contrast is the ratio of the boundary of the connected component and the edge image of the original image to the ratio of the boundary of the connected component, Its formula is: $\mathrm{<span\; class="notranslate">\; EdgeContrast</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; Border</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&cap;</span>\mathrm{<span\; class="notranslate">\; Edge</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; picture</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; Border</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span>}$ Among them, Border (CC) is the boundary of connected components, Edge (Picture) is the edge detection image of the original image, which is the union of Canny operator and Sobel operator, and its formula is: Edge(Picture)＝Canny(Picture)∪Sobel(Picture) 14. The method according to claim 11, wherein the shape regular features include connected component boundary roughness, number of holes, compactness and duty cycle, and its formula is, Connected component boundary roughness: $\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; Contour\; Roughness</span>}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; open</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; imfill</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; |</span>}$ Number of holes: Feature_CCHoles＝|imholes(CC)| Firmness: $\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; Compact</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span>}{{<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; Perimeter</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$ Duty cycle: $\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; OccupyRatio</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; BoundingBox</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}$ In the above formula, imfill (CC) is the operation of filling the internal holes of the connected components, 2×2 is the structural element of the morphological opening operation, and open() in the above formula is the morphological opening operation, which is a smoothing operation on the connected components. imholes(CC) indicates the number of holes in connected components, Area(CC) indicates the area of connected components, Perimeter(CC) indicates the perimeter of connected components, and Area(BoundingBox(CC)) indicates the area of the bounding box of connected components. 15. The method according to claim 11, wherein the stroke features include stroke average width and stroke width standard deviation, wherein, Average stroke width: Feature_Stroke_Mean=Mean(strokeWidth(skeleton(CC))) Stroke width standard deviation: $\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; Stroke</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; std</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; Deviation</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; strokeWidth</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; skeleton</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; mean</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; strokeWidth</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; skeleton</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}$ In the above formula, skeleton is a morphological skeleton algorithm, and the skeleton diagram is obtained by extracting the skeleton of the connected components, strokeWidth is the stroke width obtained for each point on the skeleton diagram, and Mean is the average value of all points on the skeleton diagram , to get the average width, and Deviation() means to find the standard deviation. 16. The method of claim 11, wherein the spatial consistency features include spatial consistency area ratios and spatial consistency boundary features, wherein, Spatial Consistency Area Ratio: $\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; AreaRatio</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; imdilate</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; picture</span>}<span\; class="notranslate">\; )</span>}$ Spatial Consistency Boundary Features: Feature_Boundary_S = Bound(imdilate(CC, 5×5)) In the above formula, imdilate is the operation of expanding the connected components, 5×5 is the structural element of the expansion operation, Area(imdilate(CC, 5×5)) indicates the area of the connected components after the expansion operation, Bound(imdilate( CC, 5×5)) represents the boundary of the connected components after the dilation operation. 17. A method for obtaining an image to be selected from an image, comprising the following steps: A. Decompose the image into connected components; B. Sending said connected components into the first stage of a cascade classifier of a group of different features cascaded according to the method of claim 1, said features being associated with the image to be selected, each cascade classification The non-selected connected components are discarded by the filter, and the connected components to be selected are output to the next classifier; and C. Combining the connected components to be selected output by the last classifier in the cascade classifier group to form the image to be selected. 18. The method of claim 17, wherein step A further comprises the steps of: A1. process described image with nonlinear Niblack thresholding method; A2. Decomposing the processed image into connected components. 19. The method according to claim 18, wherein the nonlinear Niblack thresholding method adds a statistical order filter to the background filter and the foreground filter of the standard Niblack method. 20. the method for claim 19, is characterized in that, the formula of described nonlinear Niblack thresholding method is: $\mathrm{<span\; class="notranslate">\; NLNiblack</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ></span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{<span\; class="notranslate">\; +</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>\\ <span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; <</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>\\ \mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; other</span>}\end{array}\right.$ ${\mathrm{<span\; class="notranslate">\; T</span>}}_{<span\; class="notranslate">\; \&PlusMinus;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; B</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&PlusMinus;</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; \&Center\; Dot;</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; )</span>$ ${\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; order</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; mean</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; B</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; B</span>}}<span\; class="notranslate">\; ]</span>$ ${\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; order</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; Deviation</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; ]</span>$ Where: k is set to 0.17-0.19 according to the standard Niblack method; f(x, y) is the pixel intensity at the (x, y) position of the input image; Mean(, W) is a mean filter with a window width of W; Deviation (, W) is a standard deviation filter with a window width of W; Order[, p, W] is an order statistical filter with p as the percentage and W as the width. 21. The method of claim 20, wherein the background filter

, the filter width W _B is set to 1/16 of the image width, and the percentage p1 is set to 50%; in the foreground filter

, the width W _F is 1/5 of W _B , and p2 is set to 80%. 22. The method according to claim 21, wherein each level of classifiers in the cascade classifiers of the set of different features in step B discards non-selected connected components and outputs to the next level of classifiers The method of selecting connected components includes the following steps: B1. Calculate the eigenvalues of all connected components received according to the features corresponding to each classifier; B2. comparing the eigenvalues of all connected components with the threshold intervals of the features; and B3. Discard the connected components whose eigenvalues are outside the threshold interval as non-selected connected components; output the connected components whose eigenvalues are within the threshold interval as connected components to be selected to the next-level classifier. 23. the method for claim 22, is characterized in that, the connected component of described cascade classifier group output is input a strong classifier again, and described strong classifier is the classifier that is trained by standard Adaboost method, The characteristics of the strong classifier are the same as the characteristics of the set of cascaded classifiers. 24. The method according to claim 23, wherein the strong classifier performs a linear combination of all eigenvalues of the connected components output by the cascaded classifier group and judges whether the connected components are to be selected Connected components, thereby discarding non-selected connected components and outputting connected components to be selected. 25. The method according to any one of claims 17-24, wherein the image to be selected is a text image, the connected components to be selected are text connected components, and the non-selected The connected components of are non-text connected components, and the features related to text images include: geometric features, edge contrast features, shape regular features, stroke features, and spatial consistency features. 26. The method of claim 25, wherein the geometric features include area ratios, length ratios and aspect ratios, wherein, The area ratio is the ratio of the area of the connected components to the image area, and its formula is: $\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; AreaRatio</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; picture</span>}<span\; class="notranslate">\; )</span>}$ The formula for the length ratio is: $\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; MLengthRatio</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; max</span>}<span\; class="notranslate">\; \{</span>\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; \}</span>}{\mathrm{<span\; class="notranslate">\; max</span>}<span\; class="notranslate">\; \{</span>\mathrm{<span\; class="notranslate">\; Pic</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; PIC</span>}<span\; class="notranslate">\; \}</span>}$ The formula for aspect ratio is: Feature_AspectRatio=max{w/h, h/w} In the above formula, CC represents the connected component, max{a, b} represents the larger value of a and b, w represents the width value of the bounding box of the connected component, h represents the height value of the bounding box of the connected component, PicW Indicates the width of the image, and PicH indicates the height of the image. 27. The method according to claim 25, wherein the edge contrast feature comprises edge contrast, and the edge contrast is the ratio of the boundary of the connected component and the edge image of the original image to the ratio of the boundary of the connected component, Its formula is: $\mathrm{<span\; class="notranslate">\; EdgeContrast</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; Border</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&cap;</span>\mathrm{<span\; class="notranslate">\; Edge</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; picture</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; Border</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span>}$ Among them, Border (CC) is the boundary of connected components, Edge (Picture) is the edge detection image of the original image, which is the union of Canny operator and Sobel operator, and its formula is: Edge(Picture)＝Canny(Picture)∪Sobel(Picture) 28. The method of claim 25, wherein the shape regular features include connected component boundary roughness, number of holes, compactness and duty cycle, the formula of which is, Connected component boundary roughness: $\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; Contour\; Roughness</span>}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; open</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; imfill</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; |</span>}$ Number of holes: Feature_CCHoles=|imholes(CC) compactness: $\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; Compact</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span>}{{<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; Perimeter</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$ Duty cycle: $\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; OccupyRatio</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; BoundingBox</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}$ In the above formula, imfill (CC) is the operation of filling the internal holes of the connected components, 2×2 is the structural element of the morphological opening operation, and open() in the above formula is the morphological opening operation, which is a smoothing operation on the connected components. imholes(CC) indicates the number of holes in connected components, Area(CC) indicates the area of connected components, Perimeter(CC) indicates the perimeter of connected components, and Area(BoundingBox(CC)) indicates the area of the bounding box of connected components. 29. The method according to claim 25, wherein the stroke features include stroke average width and stroke width standard deviation, wherein, Average stroke width: Feature_Stroke_Mean=Mean(strokeWidth(skeleton(CC))) Stroke width standard deviation: $\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; Stroke</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; std</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; Deviation</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; strokeWidth</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; skeleton</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; mean</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; strokeWidth</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; skeleton</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}$ In the above formula, skeleton is a morphological skeleton algorithm, which extracts the skeleton of connected components to obtain a skeleton diagram, strokeWidth is the stroke width obtained for each point on the skeleton diagram, and Mean is the average value of all points on the skeleton diagram , to get the average width, and Deviation() means to find the standard deviation. 30. The method of claim 25, wherein the spatially consistent features include spatially consistent area ratios and spatially consistent boundary features, wherein, $\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; AreaRatio</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; imdilate</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; picture</span>}<span\; class="notranslate">\; )</span>}$ Feature_Boundary_S = Bound(imdilate(CC, 5×5)) In the above formula, imdilate is the operation of expanding the connected components, 5×5 is the structural element of the expansion operation, Area(imdilate(CC, 5×5)) indicates the area of the connected components after the expansion operation, Bound(imdilate( CC, 5×5)) represents the boundary of the connected components after the dilation operation. 31. A device for acquiring an image to be selected from images, characterized in that the device comprises: decomposition means for decomposing the image into connected components; A group of cascade classifiers of different features cascaded according to the method of weight 1, the connected components are input into the first stage of the group of different features of the cascade classifiers, and each cascade classifier discards non-selected The specified connected components, and output the connected components to be selected to the next classifier; and The image synthesis device is used for combining the connected components to be selected output by the last classifier in the cascade classifier group to form the image to be selected. 32. The device of claim 31, wherein the decomposition device further comprises: Processing means, process described image with nonlinear Niblack thresholding method; Image decomposing means for decomposing the processed image into connected components. 33. The device according to claim 32, wherein the nonlinear Niblack thresholding method adds a statistical order filter to the background filter and the foreground filter of the standard Niblack method. 34. The device according to claim 33, wherein the formula of the nonlinear Niblack thresholding method is: $\mathrm{<span\; class="notranslate">\; NLNiblack</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ></span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{<span\; class="notranslate">\; +</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>\\ <span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; <</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>\\ \mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; other</span>}\end{array}\right.$ ${\mathrm{<span\; class="notranslate">\; T</span>}}_{<span\; class="notranslate">\; \&PlusMinus;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; B</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&PlusMinus;</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; \&Center\; Dot;</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; )</span>$ ${\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; order</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; mean</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; B</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; B</span>}}<span\; class="notranslate">\; ]</span>$ ${\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; order</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; Deviation</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; ]</span>$ Where: k is set to 0.17-0.19 according to the standard Niblack method; f(x, y) is the pixel intensity at the (x, y) position of the input image; Mean(, W) is a mean filter with a window width of W; Deviation (, W) is a standard deviation filter with a window width of W; Order[, p, W] is an order statistical filter with p as the percentage and W as the width. 35. The apparatus of claim 34, wherein the background filter , the filter width W _B is set to 1/16 of the image width, and the percentage p1 is set to 50%; in the foreground filter

, the width W _F is 1/5 of W _B , and p2 is set to 80%. 36. The apparatus of claim 35, wherein each stage of classifiers in the set of cascaded classifiers comprises: A computing device, configured to calculate the received eigenvalues of all connected components according to the features corresponding to each classifier; Comparing means for comparing the eigenvalues of all connected components with threshold intervals of said features, and The output device discards the connected components whose eigenvalues are outside the threshold interval as non-selected connected components; outputs the connected components whose eigenvalues are within the threshold interval as the connected components to be selected to the next classifier . 37. The device according to claim 36, wherein the connected components output by the cascaded classifier group are further input into a strong classifier, and the strong classifier is a classifier trained by a standard Adaboost method, and the The characteristics of the strong classifier are the same as the characteristics of the set of cascaded classifiers. 38. The device according to claim 37, wherein the strong classifier performs a linear combination of all eigenvalues of the connected components output by the cascaded classifier group and judges whether the connected components are to be selected Connected components, thereby discarding non-selected connected components and outputting selected connected components. 39. The device according to any one of claims 31-38, wherein the image to be selected is a text image, the connected components to be selected are text connected components, and the non-selected Connected components are non-text connected components, and features related to text images include: geometric features, edge contrast features, shape regular features, stroke features, and spatial consistency features. 40. The apparatus of claim 39, wherein the geometric features include area ratios, length ratios, and aspect ratios, wherein, The area ratio is the ratio of the area of the connected components to the image area, and its formula is: $\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; AreaRatio</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; picture</span>}<span\; class="notranslate">\; )</span>}$ The formula for the length ratio is: $\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; MLengthRatio</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; max</span>}<span\; class="notranslate">\; \{</span>\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; \}</span>}{\mathrm{<span\; class="notranslate">\; max</span>}<span\; class="notranslate">\; \{</span>\mathrm{<span\; class="notranslate">\; Pic</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; PIC</span>}<span\; class="notranslate">\; \}</span>}$ The formula for aspect ratio is: Feature_AspectRatio=max{w/h, h/w} In the above formula, CC represents the connected component, max{a, b} represents the larger value of a and b, w represents the width value of the bounding box of the connected component, h represents the height value of the bounding box of the connected component, PicW Indicates the width of the image, and PicH indicates the height of the image. 41. The device according to claim 39, wherein the edge contrast feature comprises edge contrast, and the edge contrast is the ratio of the coincidence degree of the boundary of the connected component and the edge image of the original image to the boundary of the connected component, Its formula is: $\mathrm{<span\; class="notranslate">\; EdgeContrast</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; Border</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&cap;</span>\mathrm{<span\; class="notranslate">\; Edge</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; picture</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; Border</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span>}$ Among them, Border (CC) is the boundary of connected components, Edge (Picture) is the edge detection image of the original image, which is the union of Canny operator and Sobel operator, and its formula is: Edge(Picture)＝Canny(Picture)∪Sobel(Picture) 42. The apparatus according to claim 39, wherein the shape regular features include connected component boundary roughness, number of holes, compactness and duty cycle, the formula of which is, Connected component boundary roughness: $\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; Contour\; Roughness</span>}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; open</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; imfill</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; |</span>}$ Number of holes: Feature_CCHoles=|imholes(CC) compactness: $\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; Compact</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span>}{{<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; Perimeter</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$ Duty cycle: $\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; OccupyRatio</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; BoundingBox</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}$ In the above formula, imfill (CC) is the operation of filling the internal holes of the connected components, 2×2 is the structural element of the morphological opening operation, and open() in the above formula is the morphological opening operation, which is a smoothing operation on the connected components. imholes(CC) indicates the number of holes in connected components, Area(CC) indicates the area of connected components, Perimeter(CC) indicates the perimeter of connected components, and Area(BoundingBox(CC)) indicates the area of the bounding box of connected components. 43. The device according to claim 39, wherein the stroke features include stroke average width and stroke width standard deviation, wherein, Average stroke width: Feature_Stroke_Mean=Mean(strokeWidth(skeleton(CC))) Stroke width standard deviation: $\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; Stroke</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; std</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; Deviation</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; strokeWidth</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; skeleton</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; mean</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; strokeWidth</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; skeleton</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}$ In the above formula, skeleton is a morphological skeleton algorithm, and the skeleton diagram is obtained by extracting the skeleton of the connected components, strokeWidth is the stroke width obtained for each point on the skeleton diagram, and Mean is the average value of all points on the skeleton diagram , to get the average width, and Deviation() means to find the standard deviation. 44. The apparatus of claim 39, wherein the spatially consistent features include a spatially consistent area ratio and a spatially consistent boundary feature, wherein, Spatial Consistency Area Ratio: $\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; AreaRatio</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; imdilate</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; CC</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; picture</span>}<span\; class="notranslate">\; )</span>}$ Spatial Consistency Boundary Features: Feature_Boundary_S = Bound(imdilate(CC, 5×5)) In the above formula, imdilate is the operation of expanding the connected components, 5×5 is the structural element of the expansion operation, Area(imdilate(CC, 5×5)) indicates the area of the connected components after the expansion operation, Bound(imdilate( CC, 5×5)) represents the boundary of the connected components after the dilation operation.

Images