CN104866850B - A kind of optimization method of text image binaryzation - Google Patents

Abstract

The invention discloses an optimization method for binarization of text images. The invention has the following technical effects: (1) The invention proposes a method for optimizing binarization. Existing binarization algorithms have different characteristics and accuracy for text images of different degradation types, but the present invention can perform secondary optimization on the basis of existing binary images, retaining binary images The advantages of the binarization algorithm itself further improve the accuracy of the binarization method. (2) The present invention proposes a method of judging the classification of all pixels in each region, each row or column of pixels by the percentage of a certain type of pixel as a feature, which can not only be used in this invention , for many other situations that require detailed classification, this method can be used for secondary classification when the classification information has been obtained initially.

Description

An Optimization Method for Text Image Binarization

technical field

The invention relates to a binarization optimization method of a text image, which belongs to the field of image processing.

Background technique

At present, with the development of time, there are more and more paper-based documents, which need to occupy more and more places for storage, and are inconvenient to use and search. Therefore, they need to be digitized and stored for dissemination, management and application. The digitization of text images requires steps such as segmentation and recognition of characters. Before performing these operations, it is necessary to binarize the text images. The accuracy of binarization directly affects whether the subsequent steps of analysis and recognition can be smooth. Therefore, the binarization of text images plays a crucial role. The binarization of a text image is to convert a grayscale image of a text into a binary image with only black and white, that is, the image is divided into two parts, the foreground and the background. Many text images are old documents, and it is inevitable that image degradation will occur. There are many reasons for the degradation, such as the source of image acquisition, the storage environment of the image, the storage time, etc., which will cause serious image degradation. The foreground and background in the text image are highly similar and difficult to distinguish, so how to accurately binarize the text image has always been a difficult problem.

The binarization algorithm has made great progress through the development of the past ten years. But for historical text images, usually the image degradation is serious and the image quality is poor. There are various types of degradation, such as illumination changes, stains, creases, handwriting soaked from the back, etc. The existing binarization algorithms have their own different characteristics for various types of degraded text images and Accuracy varies. Therefore, this paper hopes to improve the adaptability of the existing binarization algorithm to various types of degraded texts, and perform secondary optimization on the basis of the binary image obtained by the existing binarization algorithm to further improve the accuracy of binarization. Spend.

Contents of the invention

The purpose of the present invention is to provide an optimization method for binarizing text images.

The technical scheme of the present invention is,

An optimization method for text image binarization, comprising the following steps:

step 1:

The binary image obtained after binarizing the original text image with a binarization algorithm is used as the initial binary image;

Apply the k-means algorithm to the original text image, and classify the objects with the pixel values of all pixels in the image. The number of classifications is set to k. After classification, you can get a picture marked as {I ₁ , I ₂ , I ₃ …I _i }The image of k-type pixel sets, each pixel is marked as one of 1~k, calculate the average value A _i of all the pixel values in each I _i set, denote I _min is a set of pixels with the smallest average pixel value;

Step 2:

Use the RMharalick connected region detection algorithm to mark each independently closed connected region in the initial binary image. The connected region refers to a maximum connected subset in the image. In a maximum connected subset, any two pixel points P ₁ (x ₁ , y ₁ ) satisfy:

1≤(x ₁ -x ₂ ) ² +(y ₁ -y ₂ ) ² ≤2 (1)

Obtain an image with the same size as the initial binary image and marked as 1-m connected regions;

Step 3:

First, for the 1~m connected regions marked in step 2, any connected region marked as j (1≤j≤m) is randomly selected, the total number of pixels in the j connected region is S _j , and the I _min contained in it is calculated The total number of class pixels, denoted as m _j , the pixels in the j-connected region are reclassified by the following rules, Q _j (x) represents all the pixels in the j-connected region:

After the pixels in each connected region marked 1-m are divided by the above formula, a part of the region that is wrongly divided into foreground pixels can be removed, and the binary image B2 is obtained _;

Then, re-detect the connected regions on the binary image B ₂ to obtain the connected region graph marked 1~n, and randomly select the connected regions marked j (1≤j≤n), and each image is a 2-dimensional pixel matrix , each pixel in the matrix has its own row and column subscripts, assuming that the row subscripts of the pixels in the j-connected area are p~q, and the I _min of the fth (p≤f≤q) row in the j area is counted The number of class pixels m _jf and the total number of pixels S _jf , reclassify the pixels in the fth row according to the following formula, Q _jf (x) represents all the pixels in the f row in the connected area:

Use the above formula to judge each row of pixels in the connected region and record the category of the row of pixels and record it in the array a _jf . If the row is classified as a background pixel, the classification of the row The result is recorded as 0 and stored in the array, otherwise, the classification result is recorded as 1, and the formula is expressed as follows:

Each number in a _jf corresponds to a row of pixels in p~q, so when the pixels in a connected region are classified row by row using the above formula, a tag array a _jf can be generated as an additional condition. It is only necessary to re-mark the row of pixels represented by all 0s between 1 and 1 in all elements of the array a _jf as foreground pixels, and use the above method to divide all the row pixels in the connected region marked 1~n. Finally, re-detect the connected area, classify each column of pixels in each connected area obtained by the same method as above, and obtain a medium-term binary image;

Step 4:

Use a fusion rule proposed by Su, that is, for the initial binary image and the intermediate binary image, the pixels that are classified as foreground or background at the same time are considered to be classified correctly, and other different character pixels are classified as undetermined pixels; as shown in the following formula:

K(x) represents a pixel in the original image, and N _i (x) represents the value of the pixel at position K(x) in the initial binary image and the intermediate binary image. For pending pixels, the classification is done by the following formula:

In formula (6), J(x) represents the undetermined pixel, Con(x), I(x) represents the contrast value and pixel value of pixel J(x); Con _F , I _F represents the center of J(x) The average contrast value and the average gray value of the foreground pixels in the local window of Con _B and I _B represent the average contrast value and the average gray value of the background pixels in the local window;

In formula (7), (x, y) represents the row and column coordinates of pixels in the original image, Con(x, y) represents the contrast value of each pixel, and the gray value I(x, y) represents the pixel (x , y), take a 10×10 pixel window with (x, y) as the center point, and f _max (x, y) represents the maximum pixel value in the window. ε is a positive minimization factor, in order to prevent the denominator from being zero;

Note that the medium-term binary image is B ₃ , the initial binary image is B ₁ , select B ₁ as the initial iterative image, combine it with the intermediate binary image B ₃ and use formula (5) to classify, for all undetermined Pixel, take each undetermined pixel as the center to take a 3×3 window, when there is one or more foreground or background pixels in the window, according to the judgment condition of J(x) in the formula (6), the undetermined pixel is processed Classification, after the classification is completed, replace the next undetermined pixel to continue the classification, compare the image obtained after all undetermined pixels are classified with the first iteration image B ₁ , if they are not the same, use it as the initial image of the second iteration, and continue with The medium-term binary image B and ₃ are combined and classified, and the previous steps are repeated until it is basically unchanged, and the final optimized binary image is obtained.

The present invention has the following technical effects. (1) The present invention proposes a method for optimizing binarization. Existing binarization algorithms have different characteristics and accuracy for text images of different degradation types, but the present invention can perform secondary optimization on the basis of existing binary images, retaining binary images The advantages of the binarization algorithm itself further improve the accuracy of the binarization method. (2) The present invention proposes a method of judging the classification of all pixels in each region, each row or column of pixels by the percentage of a certain type of pixel as a feature, which can not only be used in this invention , for many other situations that require detailed classification, this method can be used for secondary classification when the classification information has been obtained initially.

Description of drawings

FIG. 1 is a specific flow chart of a text image binarization secondary optimization method.

Figure 2 is the original text image.

Figure 3 is the binarization diagram of the original text image by Otsu algorithm.

Fig. 4 is a binary map optimized by the present invention.

Figure 5 is the truth map of the binary image of the original text image.

Figure 6 is the original text image.

Fig. 7 is the binarization diagram of the original text image by Lelore algorithm.

Fig. 8 is a binary image after the original text image is optimized by the present invention.

Figure 9 is the truth map of the binary image of the original text image.

Detailed ways

As shown in Figure 1, input the existing binarization algorithm to obtain the initial binary image, and detect connected regions on it. For the original text image, the k-means algorithm is used to classify, and a class of pixels with the smallest mean value of the pixel value is marked, and based on such pixels, the overall and local pixel values of each connected region are reclassified to achieve The purpose of optimizing the binary map. The following is the execution content of the specific steps:

Step 1: Use the binary image obtained by binarizing the original text image with the existing binarization algorithm as the initial binary image. At the same time, for the original text image, use the k-means algorithm to classify and mark the pixels The class of pixels with the smallest mean value.

Apply the k-means algorithm to the original text image, classify the objects with the pixel values of all the pixels in the image, and set the number of classifications to k. After classification, you can get a picture marked as {I ₁ , I ₂ , LI _k } The image of k pixel sets, each pixel is marked as one of 1~k, calculate the average value A _i of all the pixel values in each I _i set, denote I _min is a set of pixels with the smallest average pixel value.

Step 2: For the initial binary image, use the connected region detection algorithm to mark the connected regions, and obtain a labeled map marked with m different connected regions.

Connected region detection refers to marking each independently closed connected region in the binary image. A connected region refers to a maximum connected subset in an image. There is a path completely composed of elements in this set between any two pixels in a connected set. The pixels in this set are all foreground pixels, and judging two There are two connectivity criteria for connectivity between pixels. If two pixel points P ₁ (x ₁ , y ₁ ), P ₂ (x ₂ , y ₂ ) satisfy:

1≤(x ₁ -x ₂ ) ² +(y ₁ -y ₂ ) ² ≤2 (1)

Let's say p ₁ and p ₂ are eight-adjacent, that is, a pixel has 8 neighbors in this case. If two pixel points P ₁ (x ₁ , y ₁ ), P ₂ (x ₂ , y ₂ ) satisfy:

(x ₁ -x ₂ ) ² +(y ₁ -Y ₂ ) ² ＝1 (2)

It is said that p ₁ and p ₂ are four-adjacent, that is, one pixel has four neighbors in this case, and eight-connectivity is generally used, which is more in line with people's vision. The algorithm of connected area marking has been relatively mature after decades of development, so it is only necessary to choose a fast and accurate connected area checking algorithm. We use RMharalick's classic connected area detection algorithm to check the connected area of the initial binary image. Detection, an image with the same size as the initial binary image and marked as m different connected regions is obtained.

Step 3: First, for all pixels in each independent connected region in step 2, perform an overall re-division of background pixels and foreground pixels; then, re-detect the connected region, and for each independent connected region, each row's Pixels are divided into background pixels and foreground pixels. After all the rows in the connected region are divided, an array is generated as an additional condition for the classification of all row pixels in the connected region; finally, the connected region is re-detected, and each independent connected region , all columns undergo the same steps as the above-mentioned row processing to obtain a medium-term binary image.

First, for the 1~m connected regions marked in step 2, any connected region marked as j (1≤j≤m) is randomly selected, the total number of pixels in the j connected region is S _j , and the I _min contained in it is calculated The total number of class pixels, denoted as m _j , the pixels in the j-connected region are reclassified by the following rules, Q _j (x) represents all the pixels in the j-connected region:

After the pixels in each connected region marked 1~m are divided by the above formula, a part of the region that is wrongly classified as foreground pixels can be removed, and a binary image B ₂ is obtained.

Then, re-detect the connected regions on the binary image B ₂ to obtain the connected region graph marked 1~n, and randomly select the connected regions marked j (1≤j≤n), and each image is a 2-dimensional pixel matrix , each pixel in the matrix has its own row and column subscripts, assuming that the row subscripts of the pixels in the j-connected area are p~q, and the I _min of the fth (p≤f≤q) row in the j area is counted The number of class pixels m _jf and the total number of pixels S _jf , reclassify the pixels in the fth row according to the following formula, Q _jf (x) represents all the pixels in the f row in the connected area:

Use the above formula to judge each row of pixels in the connected region and record the category of the row of pixels and record it in the array a _jf . If the row is classified as a background pixel, the classification of the row The result is recorded as 0 and stored in the array, otherwise, the classification result is recorded as 1, and the formula is expressed as follows:

Each number in a _jf corresponds to a row of pixels in p~q, so when the pixels in a connected region are classified row by row using the above formula, a tag array a _jf can be generated as an additional condition. It is only necessary to re-mark the row of pixels represented by all 0s between 1 and 1 in all elements of the array a _jf as foreground pixels, and use the above method to divide all the row pixels in the connected region marked 1~n. Finally, re-detect the connected regions, classify each column of pixels in each connected region obtained using the same method as above, and obtain a medium-term binary image.

Step 4: In the medium-term binary image, the edges of some characters may not be smooth and complete. By fusing the intermediate binary image and the initial binary image, the edges of the characters are more complete and smooth, and the final optimized binary image is obtained.

In order to make the character edges in the medium-term binary image more complete and smooth, a fusion rule proposed by Su is used, that is, for two binary images, pixels that are classified as foreground or background at the same time are considered to be classified correctly, and other different character pixels Classification as pending pixel. As shown in the following formula:

K(x) represents a pixel in the original image, B _i (x) represents the value of the pixel at position K(x) in the initial binary image and the intermediate binary image. For the undetermined pixels, the classification is carried out by the following formula (7):

In formula (8), (x, y) represents the row and column coordinates of pixels in the original image, Con(x, y) represents the contrast value of each pixel, and the gray value I(x, y) represents the pixel (x , y), take a 10×10 pixel window with (x, y) as the center point, and f _max (x, y) represents the maximum pixel value in the window. ε is a positive minimization factor, in order to prevent the denominator from being zero.

J(x) represents the undetermined pixel in formula (7), Con(x), I(x) represents the contrast value and pixel value of pixel J(x). Con _F , I _F represent the average contrast value and average gray value of the foreground pixels in the local window centered on J(x), Con _B and I _B represent the average contrast value and average gray value of the background pixels in the local window degree value. Note that the medium-term binary image is B ₃ , the initial binary image is B ₁ , select B ₁ as the initial iterative image, combine it with the intermediate-term binary image B ₃ and use the formula (6) to classify, for all undetermined Pixels, take each undetermined pixel as the center to take a 3×3 window, when there is one or more foreground or background pixels in the window, according to the judgment condition of J(x) in the formula (7), the undetermined pixel is processed Classification, after the classification is completed, replace the next undetermined pixel to continue the classification, compare the image obtained after all undetermined pixels are classified with the first iteration image B ₁ , if they are not the same, use it as the initial image of the second iteration, and continue with The medium-term binary image B and ₃ are combined and classified, and the previous steps are repeated until it is basically unchanged, and the final optimized binary image is obtained.

The present invention can improve the adaptability of the binarization algorithm to various types of images on the basis of the existing binarization, and can be applied to various existing binarization algorithms, and can effectively improve the accuracy of the binarization, It brings a secondary improvement to the binarized indicators and visual effects.

Specific embodiments

The text images in the picture library provided in the DIBCO competition are shown in Figure 2 and Figure 6, because the picture library provides the true value map of the binary image for measuring the accuracy of the binarization algorithm. Algorithms involved in the comparison include Otsu's method and Lelore's algorithm. As a classic binarization algorithm, Otsu is shown in Figure 3 and Figure 6. It not only lays the foundation for the development of binarization algorithms, but also generally serves as an algorithm for comparison in various binarization articles. The Lelore algorithm in 2013 achieved very good results in the DIBCO competition, so choosing them as the optimized experimental object can better illustrate the optimization effect of the present invention.

This embodiment is carried out on a 64-bit pc based on the windows7 system through the programming experiment environment of matlab2013. The text image in the accompanying drawing 2 is a text image in png image format with 1078×2477 pixels. The text image 2 in the accompanying drawing 6 is a text image in png image format with 2278×870 pixels. _In this embodiment, the values _of T1 and T2 are set to 0.2, the upper limit of the number of iterations in step 4 is set to 20, and the number of categories k in the k-means clustering algorithm is set to 3. First, use the otsu algorithm to binarize the text image 1 in the accompanying drawing 2, and obtain the binary image of the otsu algorithm as shown in the attached drawing 3. Then, according to the 4 steps in the specification, use the above parameters to process the binary image of the otsu algorithm See Figure 4 for the optimized binary image obtained after optimization. Similarly, use the Lelore algorithm to binarize the text image in accompanying drawing 6, and obtain the Lelore algorithm binary image in accompanying drawing 7, see accompanying drawing 8, and then, according to the 4 steps in the specification, use the above parameters to Lelore After the algorithm binary image is optimized, the optimized binary image is shown in Figure 8.

Accompanying drawing 3 is the Otsu algorithm binarization diagram of text image accompanying drawing 1, accompanying drawing 4 is the optimized binary diagram, and accompanying drawing 5 is the true value diagram of the binary diagram provided by DIBCO. It can be seen from Otsu's binary image that the Otsu algorithm, as a global threshold binarization method, is very sensitive to illumination changes, and for degraded text images with a slightly complex background, a large amount of background is retained as text information. After using the text image secondary optimization algorithm for optimization, a large amount of background noise is removed, and the accuracy of binarization is greatly improved.

The Lelore algorithm binarization map of the text image 6 is listed in the accompanying drawing 7, the optimized binary image is shown in the accompanying drawing 8, and the truth value map of the binary image provided by DIBCO is shown in the accompanying drawing 9. From the binary image of Lelore, we can see that the Lelore binarization algorithm is very accurate for the binarization of characters, but at the same time, it is too sensitive to the pixels in the background that are very similar to the characters, and it is difficult to distinguish these background pixels. , so a large number of non-character parts in the binarization result are reserved for binarization. As can be seen from the optimized binary image, the pixels other than these characters have been removed, and the characters themselves have not been affected, so the accuracy of binarization has been improved to a certain extent, especially the visual effect has been improved a lot .

From the above two experimental results, it can be seen that the present invention can effectively improve the processing effect of binarization, does not affect the character itself while removing wrong foreground pixels, and improves the accuracy of binarization while improving the accuracy of binarization. The binary image gives people a visual effect. At the same time, it can be applied to different algorithms and has good algorithm adaptability.

Claims (1)

1. an optimization method of text image binarization, it is characterized in that comprising the following steps: step 1: The binary image obtained after binarizing the original text image with a binarization algorithm is used as the initial binary image; Apply the k-means algorithm to the original text image, and classify the objects with the pixel values of all pixels in the image. The number of classifications is set to k. After classification, you can get a picture marked as {I ₁ , I ₂ , I ₃ ...I _i }The image of the k-type pixel set, each pixel is marked as one of 1~k, calculate the average value A _i of all the pixel values in each I _i set, record I _min is a set of pixels with the smallest average pixel value; Step 2: Use the RMharalick connected region detection algorithm to mark each independently closed connected region in the initial binary image. The connected region refers to a maximum connected subset in the image. In a maximum connected subset, any two pixel points P ₁ (x ₁ , y ₁ ), P ₁ (x ₂ , y ₂ ) satisfy: 1≤(x ₁ -x ₂ ) ² +(y ₁ -y ₂ ) ² ≤2 (1) Obtain an image with the same size as the initial binary image and marked as 1-m connected regions; Step 3: First, for the 1~m connected regions marked in step 2, any connected region marked as j (1≤j≤m) is randomly selected, the total number of pixels in the j connected region is S _j , and the I _min contained in it is calculated The total number of class pixels, recorded as m _j , T ₁ is a set threshold, the pixels in the j-connected region are reclassified by the following rules, Q _j (x) represents all the pixels in the j-connected region: After the pixels in each connected region marked 1-m are divided by the above formula, a part of the region that is wrongly divided into foreground pixels can be removed, and the binary image B2 is obtained _; Then, re-detect the connected regions on the binary image B ₂ to obtain the connected region graph marked 1~n, and randomly select the connected regions marked j (1≤j≤n), and each image is a 2-dimensional pixel matrix , each pixel in the matrix has its own row and column subscripts, assuming that the row subscripts of the pixels in the j-connected area are p~q, and the T _min of the f (p≤f≤q) row in the j area is counted The number of class pixels m _jf and the total number of pixels S _jf , T ₂ is a set threshold, and the pixels in row f are reclassified according to the following formula, Q _jf (x) represents all pixels in row f in the connected area: Use the above formula to judge each row of pixels in the connected region and record the category of the row of pixels and record it in the array a _jf . If the row is classified as a background pixel, the classification of the row The result is recorded as 0 and stored in the array, otherwise, the classification result is recorded as 1, and the formula is expressed as follows: Each number in a _jf corresponds to a row of pixels in p~q, so when the pixels in a connected region are classified row by row using the above formula, a tag array a _jf can be generated as an additional condition; just put In all the elements of the array _ajf , all the rows of pixels represented by 0 between 1 and 1 can be remarked as foreground pixels, and all the row pixels in the connected region marked 1~n are divided by the above method; finally, re-detect Connected area, for each connected area obtained Each column of pixels in the domain is classified by the same method as above, and a medium-term binary image is obtained; Step 4: Use a fusion rule proposed by Su, that is, for the initial binary image and the intermediate binary image, the pixels that are classified as foreground or background at the same time are considered to be classified correctly, and other different character pixels are classified as undetermined pixels; the following formula Show: K(x) represents a pixel in the original image, B _i (x) represents the value of the K(x) pixel in the initial binary image and the intermediate binary image, and the undetermined pixels are classified by the following formula: In formula (6), J(x) represents the undetermined pixel, Con(x), I(x) represents the contrast value and pixel value of pixel J(x); Con _F , I _F represents the center of J(x) The average contrast value and the average gray value of the foreground pixels in the local window of Con _B and I _B represent the average contrast value and the average gray value of the background pixels in the local window; In formula (7), (x, y) represents the row and column coordinates of pixels in the original image, Con(x, y) represents the contrast value of each pixel, and the gray value I(x, y) represents the pixel (x , y), take a 10×10 pixel window with (x, y) as the center point, f _max (x, y) represents the maximum pixel value in the window; ε is a positive minimization factor, In case the denominator is zero; Note that the medium-term binary image is B ₃ , the initial binary image is B ₁ , select B ₁ as the initial iterative image, combine it with the intermediate binary image B ₃ and use formula (5) to classify, for all undetermined Pixel, take each undetermined pixel as the center to take a 3×3 window, when there is one or more foreground or background pixels in the window, according to the judgment condition of J(x) in the formula (6), the undetermined pixel is processed Classification, after the classification is completed, replace the next undetermined pixel to continue the classification, compare the image obtained after all undetermined pixels are classified with the first iteration image B ₁ , if they are not the same, use it as the initial image of the second iteration, and continue with The medium-term binary image B ₃ is combined and classified, and the previous steps are repeated until the same, and the final optimized binary image is obtained.