CN106127198A - A kind of image character recognition method based on Multi-classifers integrated - Google Patents

Abstract

The invention provides an image text recognition method based on the integration of multiple classifiers, which converts the color image to be recognized into a grayscale image; performs binary processing on the grayscale image, and divides the image area containing text information; Each Chinese character is segmented from the entire text image; the grid feature and direction feature of each Chinese character are extracted; the minimum distance classifier is used, and the stroke density and total length feature is selected for the first layer of rough classification; the nearest neighbor classifier is used , select peripheral feature, grid feature and direction feature respectively to complete the classification matching of the second layer. The advantages of the present invention include: character recognition not only has strong anti-interference ability, but also has a strong ability to describe the local structure of characters, and is less affected by the stroke width; The combined classifier integration technology makes the system more reliable; it can intelligently recognize characters, improve the adaptability of the system and have a high recognition rate.

Description

An image text recognition method based on multi-classifier integration

technical field

The present invention relates to the field of image and character recognition, and more specifically, relates to an image and character recognition method based on multi-classifier integration.

Background technique

Social development has entered the information age. With the expansion and deepening of practical activities and the needs of socialization, human beings need to identify many types of information with complex forms and contents. People no longer stay in their own ears and eyes to directly obtain this information, but use computers to automatically input text into computers. Due to the continuous improvement of the level of science and technology, various research objects have been "imaged" and "digitized", and multimedia information based on images has quickly become an important information transmission medium. The text information in the image contains rich high-level semantic information. . Extracting these words is very helpful for the understanding, indexing and retrieval of high-level semantics of images.

Now, there are still several problems in the research of text image recognition technology. One is the large amount of image data. Generally speaking, to obtain higher recognition accuracy, the original image should have a higher resolution, at least larger than 64×64 . The second is image defacement. Due to the interference of the target environment, transmission errors, sensor errors, noise, background interference, deformation, etc., the image will be defaced. The third is accuracy, displacement, rotation, scale change, and distortion. Like human vision, there is a position change between the target and the sensor. Therefore, the system is required to be able to detect displacement, rotation, scale change, and distortion of the target correctly identify the target. The fourth is real-time performance. Most of the applications in the military field require the system to be able to identify targets in real time, which requires the system to have extremely fast output speed and identification efficiency.

In view of the current development status of character recognition systems, how to improve the recognition rate of printed characters is still a current research hotspot, and how to recognize characters in the world scene will be a direction for the development of character recognition systems. In addition, how to build a text recognition system with the characteristics of automatic layout analysis, strong fault tolerance, high recognition rate, error self-learning and self-correction, and easy expansion is the research goal of text recognition automation. Therefore, the research of image text recognition technology is particularly important.

Contents of the invention

In order to overcome at least one defect of the above-mentioned prior art, the present invention provides an automatic image character recognition method based on multi-classifier integration with high recognition rate.

In order to solve the problems of the technologies described above, the technical solution of the present invention is as follows:

A kind of image character recognition method based on multiclassifier integration, described method comprises the following steps:

S1: convert the color image to be recognized into a grayscale image, and omit this step if the image to be recognized itself is a grayscale image;

S2: Perform binarization processing on the obtained grayscale image, and segment the image area containing text information;

S3: Segment each Chinese character from the entire text image;

S4: extract the grid feature and direction feature of each Chinese character;

S5: Use the minimum distance classifier, and select the stroke density and total length feature to perform the rough classification of the first layer;

S6: The nearest neighbor classifier is used, and the combination of peripheral features, grid features and direction features is used to complete the classification matching of the second layer.

In a preferred scheme, in step S1, when converting the color image to be recognized into a grayscale image, the weighted average method is used for grayscale conversion, that is, the weighted average of the values of R, G, and B: R=G =B=a*R+b*G+c*B; wherein, R, G, and B represent red, green, and blue respectively, and a, b, and c are the weights of R, G, and B respectively, where b> a>c.

In image text recognition, the input image to be recognized is generally a color RGB image, which contains a large amount of color information. If the image is processed, the execution speed of the system will be reduced. In addition, the RGB image contains many colors that are not related to text recognition. Information is not conducive to the positioning of text, while grayscale images only contain brightness information and do not contain color information, which is conducive to further processing of images in the later stage, can improve the running speed, and is conducive to the next step of text positioning. Since the human eye is most sensitive to green, second to red, and the least sensitive to blue, under the condition of b>a>c, an easier-to-recognize grayscale image can be obtained.

In a preferred solution, in step S2, the grayscale image is binarized by using the OTSU algorithm (Otsu method or maximum inter-class variance method).

The binarization process of the image is to set the gray value of the pixel on the image to 0 or 255, that is, when all the pixels whose gray value is greater than or equal to the threshold are judged as specific objects, their gray value is 255 , otherwise, its grayscale value is 0, indicating other object areas or backgrounds, and the processed image will show an obvious black and white effect. The binarization of the image will be to divide the gray level of the pixel into 2 levels after the gray level image with 256 gray levels is selected through a suitable threshold. The properties of the image after binarization processing are only related to the position of the pixel with a gray value of 0 or 255, and no longer involve pixels of other gray levels, which is convenient for further processing of the image, and the data The amount of processing and compression is small, and the obtained binarized image can still reflect the overall and local characteristics of the image. In order to obtain an ideal binarized image, the selection of the threshold is very important. Choosing an appropriate threshold can not only effectively remove noise, but also clearly divide the image into the target area and background, greatly reducing the amount of information and improving the processing speed.

In a preferred solution, in step S3, a word segmentation method is used to identify a single character in the image area, that is, using the peaks and troughs formed by the vertical projection of the blank space between characters on the horizontal direction of the image to separate the individual characters Characters are split out.

In a preferred solution, in step S3, in order to improve the accuracy rate, the regression type character segmentation method is used to identify a single character, that is, according to the characteristics that the Chinese character is a square shape and has a roughly uniform size, it is obtained when using line segmentation The height of the text is used to estimate the width of the text, so as to predict the position of the next text.

In a preferred solution, in step S4, the specific method for extracting text grid features is as follows:

1) Divide the text lattice into 8×8 parts;

2) Find the number of black dots in each share, represented by P11, P12, ... P18, P21 ... P88;

3) Find the total number of black dots P=P11+P12+...+P18+P21+...+P88 in the text;

4) find the percentage Pij=Pij × 100/P that black dots account for the whole text black dots in every share, wherein i, j are integers greater than or equal to 1 and less than or equal to 8, feature vector (P11, P12, ... P18, P21...P88) is the grid feature of the text.

In a preferred solution, in step S4, the specific method of extracting the text direction feature is as follows:

Binarize and normalize the text bitmap image, extract the contour information, and assign one or two directions to each point on the contour. Divide the text lattice into n×n grids, calculate the number of 4 direction attributes included in each grid, thereby forming a 4-dimensional vector, and integrate all grid features to form a 4×n×n The feature vector of dimension is the direction feature.

In a preferred solution, in step S5, the specific method of constructing the minimum distance classifier is as follows:

1) Extract the stroke density length of the text from the sample as the feature vector of the rough classification. 2) Calculate the features corresponding to the samples of each category separately. Each dimension of each category has a feature set. Through the set, a mean value, that is, the feature center, can be calculated. 3) Usually, in order to eliminate the influence of different features due to different dimensions, we need to normalize the features of each dimension, or scale them to (-1,1) and other intervals to make them de-dimensionalized. 4) Use the selected distance criterion to judge the book to be classified.

In a preferred solution, in step S6, the specific method of constructing the nearest neighbor classifier is as follows:

1) Initialize the distance to the maximum value

2) Calculate the distance dist between the unknown sample and each training sample

3) Get the maximum distance maxdist among the current K nearest samples

4) If dist is less than maxdist, use the training sample as the K-nearest neighbor sample

5) Repeat steps 2, 3, and 4 until the distance between the unknown sample and all training samples is calculated

6) Count the number of occurrences of each class label in the K-nearest neighbor sample

7) Select the class label with the highest frequency as the class label of the unknown sample

Compared with the prior art, the beneficial effects of the technical solution of the present invention are: the present invention provides an image character recognition method based on multi-classifier integration, which converts the color image to be recognized into a grayscale image; value processing, and segment the image area containing text information; segment each Chinese character from the entire text image; extract the grid feature and direction feature of each Chinese character; use the minimum distance classifier, and select the total stroke density The length feature is used to carry out the rough classification of the first layer; the nearest neighbor classifier is used to combine the peripheral feature, grid feature and direction feature to complete the classification matching of the second layer. For feature extraction, the method of combining grid and direction features is adopted, so that text recognition has both strong anti-interference ability and strong ability to describe the local structure of text, and is less affected by stroke width; for image text In the identification, artificial intelligence learning technology is applied to improve the adaptability of the system and the recognition rate is high; for the design of the classifier, the classifier integration technology of the complementary combination of the minimum distance classifier and the nearest classifier is used to make the system more reliable .

Description of drawings

Fig. 1 is a flowchart of an image text recognition method based on multi-classifier integration.

Figure 2 is a schematic diagram of grayscale conversion and binarization.

Fig. 3 is a schematic diagram of the regression word segmentation method.

Figure 4 is a schematic diagram of extracting text grid features.

Fig. 5 is a schematic diagram of extracting directional grid features.

Fig. 6 is a schematic diagram of character recognition integrated with multiple classifiers.

FIG. 7 is a schematic diagram of dividing a whole paragraph of text into individual fonts.

FIG. 8 is a schematic diagram of outputting text in the form of a text box.

detailed description

The accompanying drawings are for illustrative purposes only and cannot be construed as limiting the patent;

In order to better illustrate this embodiment, some parts in the drawings will be omitted, enlarged or reduced, and do not represent the size of the actual product;

For those skilled in the art, it is understandable that some well-known structures and descriptions thereof may be omitted in the drawings.

The technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

Example 1

As shown in Figure 1, a kind of image character recognition method based on multiclassifier integration, described method comprises the following steps:

S1: convert the color image to be recognized into a grayscale image, and omit this step if the image to be recognized itself is a grayscale image;

When converting the color image to be recognized into a grayscale image, the weighted average method is used for grayscale conversion, that is, the weighted average of the values of R, G, and B: R=G=B=a*R+b*G+c *B; Among them, R, G, and B represent red, green, and blue, respectively, and a, b, and c are the weights of R, G, and B, respectively, where b>a>c.

In image text recognition, the input image to be recognized is generally a color RGB image, which contains a large amount of color information. If the image is processed, the execution speed of the system will be reduced. In addition, the RGB image contains many colors that are not related to text recognition. Information is not conducive to the positioning of text, while grayscale images only contain brightness information and do not contain color information, which is conducive to further processing of images in the later stage, can improve the running speed, and is conducive to the next step of text positioning. Since the human eye is most sensitive to green, second to red, and the least sensitive to blue, under the condition of b>a>c, an easier-to-recognize grayscale image can be obtained.

S2: Perform binarization processing on the obtained grayscale image, and segment the image area containing text information;

As shown in FIG. 2, in step S2, the grayscale image is binarized using the OTSU algorithm. The binarization process of the image is to set the gray value of the pixel on the image to 0 or 255, that is, when all the pixels whose gray value is greater than or equal to the threshold are judged as specific objects, their gray value is 255 , otherwise, its grayscale value is 0, indicating other object areas or backgrounds, and the processed image will show an obvious black and white effect. The binarization of the image will be to divide the gray level of the pixel into 2 levels after the gray level image with 256 gray levels is selected through a suitable threshold. The properties of the image after binarization processing are only related to the position of the pixel with a gray value of 0 or 255, and no longer involve pixels of other gray levels, which is convenient for further processing of the image, and the data The amount of processing and compression is small, and the obtained binarized image can still reflect the overall and local characteristics of the image. In order to obtain an ideal binarized image, the selection of the threshold is very important. Choosing an appropriate threshold can not only effectively remove noise, but also clearly divide the image into the target area and background, greatly reducing the amount of information and improving the processing speed.

The OTSU algorithm divides the image into two parts, the background and the target, according to the grayscale characteristics of the image. The larger the variance between the classes between the background and the target, the greater the difference between the two parts of the image. When part of the target is wrongly divided into background or target Misclassification of part of the background into the target will cause the difference between the two parts to become smaller. Therefore, the segmentation that maximizes the variance between classes means that the probability of misclassification is the smallest;

The steps of Otsu algorithm are as follows:

Suppose the image contains L gray levels (0,1...,L-1), the number of pixels with gray value i is Ni, and the total number of pixels in the image is N=N0+N1+...+ N(L-1), the probability of a point with a gray value i is: P(i) = N(i)/N;

Threshold t divides the entire image into two categories, dark area c1 and bright area c2, then the inter-class variance σ is a function of t: σ=a1*a2(u1-u2)^2; where aj is the value of class cj The ratio of the area to the total area of the image, a1=sum(P(i)) i->t, a2 = 1-a1;

uj is the mean value of class cj, u1 = sum(i*P(i))/a1 0->t, u2 = sum(i*P(i))/a2, t+1->L-1, this method Select the optimal threshold t^ to maximize the variance between classes, that is: let Δu=u1-u2, σb = max{a1(t)*a2(t)Δu^2}.

S3: Segment each Chinese character from the entire text image;

As shown in Figure 3, in step S3, a single character in the image area is identified using the character segmentation method, that is, a single character is segmented using the peaks and troughs formed by the vertical projection of the blank space between characters in the horizontal direction of the image come out. In order to improve the accuracy, a regression-type character segmentation method is used to identify a single character, that is, according to the characteristics of Chinese characters that are square and roughly uniform in size, the width of the character is estimated by using the height of the character obtained during line segmentation, so as to predict the next character. The position of the text.

S4: extract the grid feature and direction feature of each Chinese character;

Extracting a single type of feature for Chinese character recognition is not easy to reduce the false recognition rate, and it is not easy to improve the anti-interference ability. Because the amount of information of Chinese characters used in this way is limited, it cannot fully reflect the characteristics of Chinese characters. For any feature, there must be a "dead angle" in its recognition, that is, it is difficult to distinguish Chinese characters by using this feature. From the perspective of pattern recognition, if the space composed of all the vectorized features of Chinese characters is called space Ω (i=1, 2,...), then the information of the entire space Ω is used for Chinese character recognition, since the provided The information of Chinese characters is sufficient, and the anti-interference ability will be greatly enhanced. However, in practical applications, the trade-off between recognition accuracy, recognition speed (calculation amount) and system resources must be considered. So any practical OCR system only utilizes the information of some of the subspaces. Due to the lack of information, it is inevitable to encounter the problem of identifying "dead spots".

Based on the research of these methods, the present invention selects the grid features and direction features of Chinese characters to recognize Chinese characters. These features have strong anti-interference ability and strong ability to describe the local structure of characters. The influence of each is small, and they complement each other, so that the "dead angle" of Chinese character recognition is greatly reduced, thereby improving the recognition rate.

As shown in Figure 4, in step S4, the specific method for extracting text grid features is as follows:

1) Divide the character dot matrix into m×m parts, which are 8×8 parts in this embodiment.

2) Find the number of black dots in each share, expressed by P11, P12, ... P18, P21 ... P88.

3) Calculate the total number of black dots in the text P=P11+P12+...+P18+P21+...+P88.

4) find the percentage Pij=Pij × 100/P that black dots account for the whole text black dots in every share, wherein i, j are integers greater than or equal to 1 and less than or equal to 8, feature vector (P11, P12, ... P18, P21...P88) is the grid feature of the text.

As shown in Figure 5, in step S4, the specific method of extracting the text direction feature is as follows:

Binarize and normalize the text bitmap image, extract the contour information, and assign one or two directions to each point on the contour. Divide the text lattice into n×n grids, calculate the number of 4 direction attributes included in each grid, thereby forming a 4-dimensional vector, and integrate all grid features to form a 4×n×n The feature vector of dimension is the direction feature.

S5: As shown in Figure 6, the minimum distance classifier is used, and the stroke density and total length feature is selected to perform the rough classification of the first layer;

The minimum distance classifier uses the stroke density and total length feature to perform the rough classification of the first layer. In this method, the distance between the recognized pattern and the samples belonging to the pattern category is the smallest. Assume that the eigenvectors of c categories representing patterns are denoted by R1,...,Rc, x is the eigenvector of the recognized pattern, |x-Ri| is the distance between x and Ri (i=1, 2,...,c) , if |x-Ri| is the smallest, then divide x into the i-th category.

S6: The nearest neighbor classifier is used, and the combination of peripheral features, grid features and direction features is used to complete the classification matching of the second layer.

The nearest neighbor classifier selects the combination of grid features and direction features to complete the classification matching of the second layer. The nearest neighbor classifier expands on the basis of the minimum distance classification, uses each sample in the training set as the basis for discrimination, and searches for the samples in the training set that are closest to the sample to be classified, and uses this as the basis for classification.

After many experiments and researches, the conclusion shows that the principle of a single recognizer cannot fundamentally improve the performance of the system, and it should rely on the integration of the recognition results of multiple classifiers. Multi-classifier integration is to improve the performance of a single classifier through multiple complementary classifiers, and obtain a more reliable recognition system. Therefore, the present invention adopts the integration of the minimum distance classifier and the nearest neighbor classifier, and further improves the recognizability and accuracy of characters by optimizing the classifier design.

In order to verify the effectiveness of the present invention, relevant experiments need to be carried out. The present invention uses an original image containing 697 Chinese characters for testing. First, convert the original image into a grayscale image for the next step. The whole paragraph of text is divided into individual fonts through the regression character segmentation method. The test effect is shown in Figure 7, and each Chinese character can be accurately segmented. Finally, the method of multi-feature extraction and multi-classifier integration is used to recognize the segmented text and output it in the form of a text box. The test results are shown in Figure 8, and the results are all correct.

The multi-feature extraction method and the multi-classifier integration method make it possible to improve the recognition rate of image characters, and its good recognition effect has attracted people's attention and has broad application prospects. The invention realizes image text recognition based on a multi-classifier integration method, and makes processing and extraction of image text information more feasible.

Apparently, the above-mentioned embodiments of the present invention are only examples for clearly illustrating the present invention, rather than limiting the implementation of the present invention. For those of ordinary skill in the art, other changes or changes in different forms can be made on the basis of the above description. It is not necessary and impossible to exhaustively list all the implementation manners here. All modifications, equivalent replacements and improvements made within the spirit and principles of the present invention shall be included within the protection scope of the claims of the present invention.

Claims (7)

1. a kind of image text recognition method based on multiclassifier integration, it is characterized in that, described method comprises the following steps: S1: convert the color image to be recognized into a grayscale image, and omit this step if the image to be recognized itself is a grayscale image; S2: Perform binarization processing on the obtained grayscale image, and segment the image area containing text information; S3: Segment each Chinese character from the entire text image; S4: extract the grid feature and direction feature of each Chinese character; S5: Use the minimum distance classifier, and select the stroke density and total length feature to perform the rough classification of the first layer; S6: The nearest neighbor classifier is used, and the combination of peripheral features, grid features and direction features is used to complete the classification matching of the second layer. 2. the image character recognition method based on multi-classifier integration according to claim 1, is characterized in that, in step S1, when the colored image to be recognized is converted into a grayscale image, the weighted average method is used to carry out the grayscale conversion , that is, the weighted average of the values of R, G, and B: R=G=B=a*R+b*G+c*B; wherein, R, G, and B represent red, green, and blue, respectively, and a, b , c are the weights of R, G, and B respectively, where b>a>c. 3. the image character recognition method based on multiclassifier integration according to claim 1, is characterized in that, in step S2, adopts OTSU algorithm to carry out binarization process to grayscale image, OTSU algorithm is to press the grayscale characteristic of image , Divide the image into background and target. The larger the inter-class variance between the background and the target, the greater the difference between the two parts that make up the image. When part of the target is wrongly divided into the background or part of the background is wrongly divided into the target, it will cause 2 Part of the difference becomes smaller, so the split that maximizes the variance between classes means the smallest probability of misclassification; The steps of Otsu algorithm are as follows: Suppose the image contains L gray levels (0,1...,L-1), the number of pixels with gray value i is Ni, and the total number of pixels in the image is N=N0+N1+...+ N(L-1), the probability of a point with a gray value i is: P(i) = N(i)/N; Threshold t divides the entire image into two categories, dark area c1 and bright area c2, then the inter-class variance σ is a function of t: σ=a1*a2(u1-u2)^2; where aj is the value of class cj The ratio of the area to the total area of the image, a1=sum(P(i)) i->t, a2 = 1-a1; uj is the mean value of class cj, u1 = sum(i*P(i))/a1 0->t, u2 = sum(i*P(i))/a2, t+1->L-1, this method Select the optimal threshold t^ to maximize the variance between classes, that is: let Δu=u1-u2, σb = max{a1(t)*a2(t)Δu^2}. 4. the image text recognition method based on multi-classifier integration according to claim 1, is characterized in that, in step S3, adopts word segmentation method to identify the single text in the image area, promptly utilizes the blank between word and word The peaks and troughs formed by the vertical projection of the gap on the horizontal direction of the image separate individual characters. 5. the image text recognition method based on multi-classifier integration according to claim 4, is characterized in that, in step S3, adopts the regression type character segmentation method to identify single text, promptly according to Chinese character is a square figure, has roughly uniform The characteristics of size, use the height of the text obtained during line segmentation to estimate the width of the text, so as to predict the position of the next text. 6. the image text recognition method based on multiclassifier integration according to claim 1, is characterized in that, in step S4, the concrete method of extracting text grid feature is as follows: 1) Divide the text lattice into 8×8 parts; 2) Find the number of black dots in each share, represented by P11, P12, ... P18, P21 ... P88; 3) Find the total number of black dots P=P11+P12+...+P18+P21+...+P88 in the text; 4) find the percentage Pij=Pij × 100/P that black dots account for the whole text black dots in every share, wherein i, j are integers greater than or equal to 1 and less than or equal to 8, feature vector (P11, P12, ... P18, P21...P88) is the grid feature of the text. 7. the image text recognition method based on multiclassifier integration according to claim 1, is characterized in that, in step S4, the concrete method of extracting text direction feature is as follows: Binarize and normalize the text bitmap image, extract the contour information, and assign one or two directions to each point on the contour. Divide the text lattice into n×n grids, calculate the number of 4 direction attributes included in each grid, thereby forming a 4-dimensional vector, and integrate all grid features to form a 4×n×n The feature vector of dimension is the direction feature.