CN101315664A - Text Image Preprocessing Method for Text Recognition - Google Patents

Abstract

The invention is a text image preprocessing method for character recognition. It includes two steps of geometric correction of text image and binarization of dynamic domain value. The geometric correction method performs area correction on the geometric distortion of the text area of the text image captured by the camera to obtain a correction result of a rectangular area; the geometric correction method includes automatic area recognition and manual area designation. The dynamic threshold binarization method adaptively separates the text foreground and the background with uneven brightness in the text picture picked up by the camera. The feature of this method is that it can perform rectangular correction according to the recognized text area, and can be combined with any image blurring algorithm, and use the source image as a reference image to separate the text foreground from the background with uneven brightness. The resulting image obtained by the method can be used in fields such as computer character recognition, machine vision and machine understanding. The invention has the excellent effects of strong adaptability to the environment and stable and reliable algorithm.

Description

Text Image Preprocessing Method for Text Recognition

technical field

The invention relates to the technical field of image processing, in particular to a text image preprocessing method for character recognition.

Background technique

In the application of text recognition, computer vision, and machine understanding, it is one of the technical difficulties to separate and recognize text content in complex environment images. Machine understanding of text is based on text recognition results, and the better the quality of the source image used for text recognition, the higher the recognition rate. The purpose of the text image preprocessing method is to provide a binary image with small geometric distortion and perfect separation of the foreground text for the text recognition algorithm. The image preprocessing method for text recognition is generally to identify the text content area in the image, and then use image translation, rotation and stretching methods to correct the text content area, and then perform a binary value of the global fixed domain value on the correction result. processing.

At present, the image geometric correction methods for text content mainly focus on the tilt correction of text images, and correct the text to be horizontal or vertical according to the direction of the text and the space between lines. However, this skew correction method is only effective for skewed rectangular text. For the image obtained by the actual camera, due to the perspective relationship and lens nonlinear distortion, the original rectangular text area will be distorted into a trapezoid or a curved quadrilateral. At present, there is no convenient geometric correction method for distorting text and pictures into trapezoids or curved quadrilaterals. For the binarization of text images with changing brightness background, the disadvantage of using the global fixed threshold method is that it cannot achieve good results, while the local binarization threshold method for block images will lead to block boundary effects. In the field of text recognition preprocessing, there is currently no practical software that can perform effective geometric correction and text foreground extraction on text pictures taken by cameras.

Contents of the invention

The purpose of the present invention is to provide a convenient and practical text image preprocessing method for text recognition and text image preprocessing method for text recognition and machine understanding. Using this method, the text area can be identified according to the text picture taken by the camera. , and restore the recognition area to a rectangular area, realize the separation between the foreground of the text and the background of the brightness change through the binarization method of the adaptive threshold value, and provide a good image source for the text recognition method.

The main technical problem that the present invention solves is to adopt the following technical solutions to realize:

The geometric distortion of the text area of the text image captured by the camera is corrected to obtain the correction result of the rectangular area; and the text foreground and the background of uneven brightness in the text image captured by the camera are adaptively separated.

The method of area correction for the geometric distortion of the text area of the text image captured by the camera is: firstly, automatically identify or specify the area of the text image with geometric distortion captured by the camera, obtain the boundary position of the text area, and then use the text image text arrangement Inherent rectangular region characteristics, perform geometric correction on the identified or specified text region, and restore it to a rectangular text region; for any pixel in the text image, according to its relationship with the boundary position of the specified text region, find Get its corresponding position in the set rectangular area, so as to obtain the original position of the pixel in the corrected image. The color value or brightness value of all pixels in the corrected image is determined by any image interpolation algorithm with reference to the original position of the pixel in the corresponding source text image; the specific steps are as follows:

1) Carry out text area recognition to the image that is originally a rectangular text area, and obtain the boundary of the text;

2) Taking the area recognition result as a benchmark, calculate the positional relationship of the source image relative to the benchmark;

3) Determine the color value or brightness value of the corresponding pixel of the corrected image through any image interpolation algorithm, so as to obtain the result of geometric correction;

4) Step 1, step 2 and step 3 can also be repeated for the geometric correction result obtained in step 3, and a better geometric correction result can be obtained through iteration.

The text image captured by the camera or the result image obtained by the above-mentioned text area geometric distortion correction method is used to extract the text foreground, and the method for adaptively separating the text foreground and the background with uneven brightness is image dynamic domain binary value Preprocessing method: Through the operation between any pixel in the image and its neighboring pixels, any image blurring method can be used, but not limited to, to obtain the local background brightness; by using the weighted value of the local background brightness as the area The binarization threshold is used to separate the text foreground of the area; according to the separation result, the text foreground part in the source image is deducted, and the local background brightness is calculated by using but not limited to any image blurring method, and then its weighted value is used As a new dynamic threshold, the source image is binarized to obtain a more accurate separation result; this process can be iterated multiple times, and the specific steps are as follows:

1) Use the fixed domain value method to roughly separate the text foreground;

2) Use any image blurring method to calculate the local background brightness of the roughly separated background area;

3) Carry out weighted calculation on the local background brightness, and use it as the dynamic domain value to separate the document foreground of the original image;

4) Using the result obtained in step 3, iterative steps 2 and 3 are repeated to obtain more accurate separation results.

Compared with the prior art, the method of the present invention has the advantages that: because the present invention adopts the above-mentioned technical measures, compared with the prior art method, it has the excellent effects of strong adaptability to the photographic environment, simple algorithm, stability and reliability, and in addition , an iterative method is also designed to improve the quality of the resulting image, which has a wider range of applications. In practice, the text image preprocessing effect using this method is obviously better than the result of the existing text recognition preprocessing method. The present invention can adopt software to realize, and form the software part or module that computer text recognition software carries out image preprocessing, also can realize through hardware or digital signal processing chip, become a function of embedded systems such as digital camera, video camera, robot vision system .

The specific implementation method of the present invention is given in detail by the following examples and accompanying drawings.

Description of drawings

FIG. 1 is a schematic diagram of a text area of a text image corrected for linear distortion (the corresponding undistorted text area of this area is a rectangle).

Among them: P11, P1n, Pm1, Pmn are four vertices of the trapezoidal area respectively, P1k, P21, Pin, ij are pixels at different positions in the image area.

FIG. 2 is a schematic diagram of the correction result area of the text image in FIG. 1 .

Among them: P11', P1n', Pm1', Pmn' are the correction position results corresponding to the four vertices of P11, P1n, Pm1, Pmn in Figure 1; P1k', P21', Pin', Pij' correspond to P1k in Figure 1 , P21, Pin, Pij points.

FIG. 3 is a schematic diagram of a text area of a corrected text image in the case of nonlinear distortion (the corresponding undistorted text area of this area is a rectangle).

Wherein, Pij represents a pixel in the distorted area.

FIG. 4 is a schematic diagram of a text area after vertical correction.

The pixel point Pij' corresponds to the pixel point Pij in Fig. 3 .

FIG. 5 is a schematic diagram of the text area after the area in FIG. 4 has been further corrected horizontally.

The pixel point Pij" corresponds to the pixel point Pij' in Fig. 4, that is, corresponds to the pixel point Pij in Fig. 3 .

Fig. 6 is a one-dimensional representation of a schematic diagram of a dynamic threshold binarization calculation method.

Among them, the curve f represents the two-dimensional image f(x, y); the curve g represents the blurred image g(x, y) obtained through calculation between adjacent pixels, and the curve t represents the blurred image g(x, y) after translation weighting The obtained image t(x, y), D is a translation weighted value.

Detailed ways

The text image preprocessing method of the present invention will be further described in detail below with reference to the embodiments of the accompanying drawings.

Example

1. Correction of irregular quadrilaterals for linear geometric distortion

As shown in Figure 1, the text area in the corrected image is linearly distorted, and the text part that was originally a rectangular area is distorted into a trapezoidal area. Since the part of the text area is obviously different from the edge of the text area, the area where the text is located can be automatically identified by an automatic recognition method, or the area can be manually set by manual observation.

In order to correct this distorted area to the undistorted rectangular area shown in Figure 2, the method of applying the present invention is:

Divide the irregular quadrilateral edge into several pixels horizontally and vertically. For example, count the number of pixels traversed by the line segment P11P1n, set N1, count the number of pixels traversed by the line segment Pm1Pmn, set N2, then divide the number of pixels horizontally n is

n=(N1+N2)/2

Use a similar method to divide the vertical pixel points of the line segment P11Pm1 and line segment P1nPmn, and set the number of vertically divided pixel points as m.

Calculate the coordinates of n pixel points (P11, P12, .

These points are then mapped to n points (P11', P12', ..., P1n') on the line segment P11'P1n' in Figure 2.

Operate the n pixels on the next line segment P21P2n in the same way.

In this way, when performing the i-th line segment division, the n pixel points (Pij, j=1, . Pij', j=1, . . . , n) position.

Such a process has been carried out until the line segment Pm1Pmn.

The above process can be repeated to form an iteration.

2. Correction of irregular quadrilaterals for nonlinear geometric distortion

As shown in Figure 3, nonlinear distortion occurs in the text area in the corrected image, and the text part that was originally a rectangular area is distorted into an irregular area surrounded by four curves. Since the part of the text area is obviously different from the edge of the text area, the area where the text is located can be automatically identified by an automatic recognition method, or the area can be manually set by manual observation. Due to the smooth nature of the curve, it is only necessary to determine a few points on the curve during area identification or setting, and then calculate an approximate curve through interpolation methods such as spline interpolation.

In order to correct the distortion area surrounded by this curve into the distortion-free rectangular area shown in Figure 6, the method of applying the present invention is:

An iterative correction process is divided into two stages: the first stage is to perform vertical correction, and the result is to correct the vertical perimeter of the area into two vertically equal sides, but there is still distortion in the horizontal direction, as shown in the figure 4. The second stage is to correct the horizontal direction on the basis of the results shown in Figure 4, and obtain the rectangular area correction results shown in Figure 5.

The correction algorithm used in both stages is the same. Taking the first stage of vertical correction as an example, after setting the calculation resolution m in the vertical direction, divide the two curved sides in the vertical direction into m segments, and connect the corresponding dividing points on the two curved sides with a straight line, and then according to Set the resolution n in the horizontal direction, divide the line into n segments, and then obtain m*n intersection points of these grids, and the point Pij shown in Figure 3 represents any one of these grid intersection points.

According to the serial number (i, j) of the point Pij, it can be directly mapped to the area shown in Fig. 4 to obtain the point Pij'. Such a process has been carried out until i=m, j=n.

Use the same method to complete the second stage of horizontal correction. The correction results shown in Figure 5 are obtained. Finally, the point Pij on the corrected image is mapped to the point Pij" shown in Fig. 5 .

The above process can be repeated to form an iteration.

3. Binarization separation method of text part of dynamic domain value

As shown in FIG. 6 , the brightness change of the two-dimensional image is illustrated by a one-dimensional curve in the figure. Let the part of the curve with a large value (convex part) be the text part in the image, and the part of the curve with a small value (concave part) be the background part of the image. Note that the value of the background part changes, indicating that the brightness of the image background changes with different regions. f(x, y) represents the source image. The idea of the dynamic threshold binarization separation method is to construct a quantitative threshold value that changes with the local background brightness of the image, so as to separate the raised part of the text from the background part. For this reason, the method of the present invention is:

Any form of fuzzy operation is performed on the source image f(x, y), that is, low-pass filtering of the image, and the obtained blurred image g(x, y) represents the brightness of the local area of the source image. The present invention utilizes the weighted value of the fuzzy image g(x, y)

t(x,y)=g(x,y)+D

As a dynamic domain value, binary separation is performed on the source image f(x, y), namely

H(x, y)=255 when f(x, y)>t(x, y)

H(x, y)＝0 When f(x, y)＜t(x, y)

The iterative method can also be further used to improve the separation effect of the text part, as follows:

According to the separation result, the text part in the source image f(x, y) is deducted (H(x, y)=255 parts), that is, the corresponding text part is replaced by the adjacent background brightness, and the background part f'(x, y), for which the above-mentioned image blurring method can be used to obtain a new blurred image g'(x, y), which can be used as a new dynamic domain value to perform binarization on the source image after weighting.

4. Iteration stop condition

In the above methods (geometric correction method and dynamic threshold method), when the mean square value of the difference between the results of the previous two iterations is less than the set threshold, or when the number of iterations is greater than the specified maximum number of iterations, the iteration is stopped.

The above is only a preferred embodiment of the present invention, and does not constitute any formal limitation to the present invention. Any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention, All still belong to the scope of the technical solution of the present invention.

Claims (4)

1. A text image preprocessing method for character recognition, characterized in that: carry out regional correction to the geometric distortion of the text area of the text picture taken by the camera, obtain the correction result of the rectangular area; The foreground of the text and the background with uneven brightness are adaptively separated. 2. the text image preprocessing method that is used for character recognition according to claim 1 is characterized in that: the method that area correction is carried out to the text area geometric distortion of the text picture of camera pick-up is: at first to camera pick-up has geometric distortion Automatically identify or specify the area of the text image, obtain the boundary position of the text area, and then use the inherent rectangular area characteristics of the text image text arrangement to perform geometric correction on the identified or specified text area, and restore it to a rectangular text area; For any pixel in the text image, calculate its corresponding position in the set rectangular area according to the relationship between it and the boundary position of the specified text area, so as to obtain the original position of the pixel in the corrected image. The color value or brightness value of all pixels in the corrected image is determined by any image interpolation algorithm with reference to the original position of the pixel in the corresponding source text image; the specific steps are as follows: 1) Carry out text area recognition to the image that is originally a rectangular text area, and obtain the boundary of the text; 2) Taking the area recognition result as a benchmark, calculate the positional relationship of the source image relative to the benchmark; 3) Determine the color value or brightness value of the corresponding pixel of the corrected image through any image interpolation algorithm, so as to obtain the result of geometric correction; 4) Step 1, step 2 and step 3 can also be repeated for the geometric correction result obtained in step 3, and a better geometric correction result can be obtained through iteration. 3. the text image preprocessing method that is used for text recognition according to claim 1, is characterized in that: to the text picture that camera picks up or the result image that draws by the text area geometric distortion correction method described in claim 1 The method of extracting the text foreground and adaptively separating the text foreground from the background with uneven brightness is the image dynamic domain value binarization preprocessing method: through the operation between any pixel in the image and its neighboring pixels, you can use But it is not limited to use any image blurring method to obtain the local background brightness; by using the weighted value of the local background brightness as the binarization threshold of the region, the text foreground of the region is separated; according to the separation result, the source image is divided into Subtract the foreground part of the text, use but not limited to any image blurring method to calculate the local background brightness, and then use its weighted value as a new dynamic threshold to binarize the source image to obtain a more accurate separation result; this A process can be iterated multiple times, the specific steps are as follows: 1) Use the fixed domain value method to roughly separate the text foreground; 2) Use any image blurring method to calculate the local background brightness of the roughly separated background area; 3) Carry out weighted calculation on the local background brightness, and use it as the dynamic domain value to separate the document foreground of the original image; 4) Using the result obtained in step 3, iterative steps 2 and 3 are repeated to obtain more accurate separation results. 4. The text image preprocessing method for character recognition according to claim 1, characterized in that the geometric correction method includes automatic area recognition and manual area designation.

Images