CN105447489A - Character and background adhesion noise elimination method for image OCR system - Google Patents

Abstract

The invention discloses a character and background adhesion noise elimination method for an image OCR system, comprising the following steps: (1) setting the effective range of the stroke length of a character to be identified; (2) selecting an image containing the character to be identified in a natural environment, and calculating the difference boundary diagram of the image by means of mean difference template; (3) thresholding the obtained difference boundary diagram of the image to be identified to form a ternary-valued boundary diagram of the image to be identified; and (4) carrying out searching in the ternary-valued boundary diagram according to the angle-of-inclination range and the linear boundary. By using the character and background adhesion noise elimination method for the image OCR system of the invention, the problem that a character with adhesion noise is hard to position for the existing OCR application system is solved in a complex and changeable natural environment, the character position can be extracted accurately, and the real-time requirements of engineering application can be satisfied.

Description

A method for eliminating the adhesion noise between characters and background in an image OCR recognition system

technical field

The invention relates to a natural scene picture OCR recognition system, in particular to a method for eliminating the adhesion noise between characters and background in the picture OCR recognition system.

Background technique

In the field of artificial intelligence, optical character recognition (OCR) is a very important technology. Due to the popularity of smart phones and the need for classification and search of massive images in cloud storage, OCR recognition of natural scene images has become a hot research topic in recent years. OCR Technology generally includes several processes such as character positioning, character cutting, and character recognition, among which the speed and accuracy of character positioning directly affect the quality of OCR recognition technology and are the key to the entire OCR system.

OCR character positioning usually uses the structural information of the string to locate the area where the string is located through a global search, and the most common and effective method for extracting the string structure information is to extract the string image through binarization technology The commonly used binarization methods include fixed threshold method, adaptive threshold method, global threshold method and local threshold method. Various noises will be introduced more or less in complicated situations, which will easily lead to positioning failure, or generate a large amount of false character information, thus greatly increasing the calculation amount of subsequent processing.

For the noise introduced by binarization, the existing method usually uses some filtering methods to filter out the noise, such as median filtering, mathematical morphology, two-dimensional wavelet analysis, etc. These different methods have different effects on different images. Effect, although the commonly used linear low-pass filter and neighborhood averaging method can remove part of the noise, but they have the negative effect of image blur, the median filtering method can eliminate isolated noise points, and the blurring produced is relatively small, but It is not effective in removing noise from binary images. Mathematical morphology can corrode some black patches to a certain extent, but this often leads to aggravated deformation of the original image, which is not conducive to subsequent calculations.

In addition, in the OCR recognition of natural scene pictures, due to the influence of factors such as light, weather, and sundries, there are more binarized noises. In practical engineering applications, independent and discrete noises are often easier to distinguish, while characters or It is often difficult to deal with the noise caused by the adhesion of the objects to be recognized. The result of the adhesion noise is that the corresponding character position range is disturbed, resulting in the failure of accurate positioning of the characters, which affects the overall recognition rate of the system. How to eliminate the noise in natural scenes? Character adhesion noise has become a key problem in OCR recognition. Therefore, it is very important to propose a real-time character adhesion noise elimination technology that can meet the needs of practical applications and realize various complex environments. The interference of the character area, thus improving the accuracy of OCR character positioning.

Contents of the invention

The purpose of the present invention is to provide a method for eliminating the cohesive noise between characters and the background of a picture OCR recognition system. Aiming at the OCR technology of existing natural scene pictures, there are deficiencies in noise processing, and a new method for eliminating cohesive noise is proposed. In the complex and changeable natural environment, this method solves the problem that the existing OCR application system has difficulty in locating characters with cohesive noise, realizes accurate extraction of character positions, and can meet the real-time requirements of engineering applications.

In order to achieve the above object, the technical solution provided by the present invention is: the character and background cohesion noise elimination method of the natural scene picture OCR recognition system includes the following steps executed in order:

A character and background cohesion noise elimination method of a picture OCR recognition system, the character and background cohesion noise elimination method of the picture OCR recognition system comprises the following steps:

1) According to the prior knowledge of the engineering environment applied by the OCR system, the effective range of the stroke length StrokeLen of the character to be recognized is set;

2) Select an image containing a character to be recognized in the natural environment, and calculate the image difference boundary map in the form of a mean difference template, and the difference boundary map includes a character area and a background area;

3) Thresholding the differential boundary map of the image to be recognized obtained above to form a three-valued boundary map of the image to be recognized, the three-valued boundary map includes a region to be recognized and a background region;

4) In the ternary boundary map of the image to be recognized obtained above, search according to the slope angle range and the linear boundary, and check whether the boundary exceeds the valid range of the stroke length. If it does not exceed the valid range of the stroke length, Then it is considered as a valid boundary point and is reserved. When it exceeds the effective range of the stroke length, the boundary is judged as noise, and it is removed from the ternary boundary image.

In the step 3), the formula for thresholding the differential boundary map of the image to be recognized is as follows: $Thre\mathrm{the\; s}hold\left(P\right(x,\mathrm{the\; y}\left)\right)=\left\{\begin{array}{cc}c1,& ifP(x,\mathrm{the\; y})<ParamA\\ c2,& ifP(x,\mathrm{the\; y})>ParamB\\ c3,& other\mathrm{the\; s}\end{array}\right.$

Among them, P(x, y) is the difference value of the pixel point (x, y) in the difference map, ParamA and ParamB are preset thresholds, c1, c2 and c3 are unsigned integers, and the value range is [0, 255 ]; where c1 represents the border where the brightness of the current pixel is brighter than the surrounding, c2 represents the border where the brightness of the current pixel is darker than the surrounding, and c3 represents the value of the non-border.

The color of c1 is black and its value is c1=0; the color of c2 is white and its value is c2=255; the color of c3 is gray and its value is c3=128; The values are 0 and 255 respectively, and the value of the non-boundary pixels is 128, that is, the black and white pixels are boundary pixels, and the gray pixels are non-boundary pixels.

In the step 4), the algorithm steps of searching the three-valued image of the image to be recognized according to the linear boundary in the range of oblique angles are as follows:

4.1) Calculate the linear tilt offset detection template with the tilt angle in the range of [α, β];

4.2) For each boundary point in the three-valued image, use the straight line detection template obtained in step a) to check the continuous boundary line starting from the boundary point, having the same boundary value and matching the template;

4.3) For the matched continuous boundary line, check whether its length belongs to the valid range, if it belongs to the valid range, keep it, otherwise the boundary line is a noise boundary, and set the boundary pixel values contained in all boundary points on the boundary line to c3 .

In the step 4-1), the algorithm steps for calculating the straight line detection template with an inclination angle in the range of [α, β] are as follows:

4.1.1) According to the longest stroke length StrokeLen set in step 1), calculate the number of straight-line templates Num, the length Len of the straight-line template lines and the maximum offset value MaxOffY of the template in the y direction according to the actual situation of the project. The calculation formulas are respectively as follows;

$\left\{\begin{array}{c}\mathrm{<span\; class="notranslate">\; N</span>}\mathrm{<span\; class="notranslate">\; u</span>}\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 4</span>}\mathrm{<span\; class="notranslate">\; S</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; r</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; k</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; no</span>}\\ \mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; S</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; r</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; k</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; no</span>}\\ \mathrm{<span\; class="notranslate">\; m</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; x</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; no</span>}\end{array}\right.$

4.1.2) For each linear tilt offset template, calculate the offset value in the y direction according to the following formula.

$\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; *</span>\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; Y</span>}}{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; \&Element;</span><span\; class="notranslate">\; \&lsqb;</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; \&rsqb;</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; \&Element;</span><span\; class="notranslate">\; \&lsqb;</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; m</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; x</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; m</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; x</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; \&rsqb;</span>$

In the step 4.2), the three-valued image of the image to be recognized is subjected to a linear boundary search algorithm, and the method of judging a continuous boundary line starting from a certain boundary point, having the same boundary value, and matching the straight-line tilt offset template is as follows :

4.2.1) When the line width of the straight line is an ordinary straight line of 1, the method starts from the starting point and checks from left to right whether there are directly adjacent equivalent boundary points, and if there is, all boundary points on the straight line are It belongs to the matching point, otherwise it belongs to the non-matching point;

4.2.2) When the line width of the straight line is N, based on the eight-connected generalized straight line, the method starts from the starting point and detects the equivalent boundary point under the eight-connected sense. When there is such a generalized straight line, then on the generalized straight line All boundary points belong to matching points, otherwise they belong to non-matching points.

The beneficial effects of the present invention are: the character and background adhesion noise elimination method of a picture OCR recognition system of the present invention proposes a new adhesion Noise elimination method, in the complex and changeable natural environment, this method solves the problem that the existing OCR application system has difficulty in locating characters with cohesive noise, realizes accurate extraction of character positions, and can meet the real-time requirements of engineering applications.

Description of drawings

Fig. 1 is the flow chart of the character and background adhesion noise elimination method of a kind of picture OCR recognition system of the embodiment of the present invention;

Fig. 2 is the mean difference template of the embodiment of the present invention;

Fig. 3 is a schematic diagram of a generalized straight line of a scanning differential boundary according to an embodiment of the present invention, the blank rectangular blocks in the figure represent non-boundary pixels, and the non-blank rectangular blocks in the figure represent boundary pixels;

Fig. 4 is a flow chart of searching for a straight-line generalized straight-line boundary on the three-valued image according to an embodiment of the present invention;

Fig. 5 is the flow chart of the straight line detection template for calculating the tilt angle in the range of [α, β] according to the present invention;

FIG. 6 is a schematic diagram of an ordinary straight line composed of differential boundary pixels in an embodiment of the present invention;

Fig. 7 is a schematic diagram of adjacent pixels based on the eight-connected meaning used in the generalized straight line judgment of the present invention;

Fig. 8 is a schematic diagram of a generalized straight line with a width of 2 in the present invention;

Fig. 9 is a schematic diagram of a generalized straight line with a width of 3 in the present invention;

Fig. 10 is the picture before the embodiment of the present invention is applied to the first license plate with adhesion;

Fig. 11 is an effect diagram after the embodiment of the present invention is applied to the first license plate processing;

Fig. 12 is the picture before the embodiment of the present invention is applied to the second license plate with adhesion;

Fig. 13 is an effect diagram after the embodiment of the present invention is applied to the second license plate processing;

Fig. 14 is the picture before the embodiment of the present invention is applied to the third license plate with adhesion;

Fig. 15 is the effect diagram after the embodiment of the present invention is applied to the processing of the 3rd license plate;

Fig. 16 is the picture before the embodiment of the present invention is applied to the 4th license plate with adhesion;

Fig. 17 is the effect diagram after the embodiment of the present invention is applied to the processing of the 4th license plate;

Fig. 18 is the picture before the embodiment of the present invention is applied to the fifth license plate with adhesion;

Fig. 19 is an effect diagram after the embodiment of the present invention is applied to the processing of the fifth license plate.

detailed description

The present invention will be described in further detail below in conjunction with the accompanying drawings.

As shown in Figure 1, it is a flow chart of a method for eliminating the adhesion noise between characters and background of a picture OCR recognition system according to an embodiment of the present invention, and the method includes:

1) According to the prior knowledge of the engineering environment applied by the OCR system, the effective range of the stroke length StrokeLen of the character to be recognized is set. This parameter is closely related to the actual project, and should be set as a configurable parameter in the actual application process. When this method is applied to the OCR recognition of bayonet license plate images, this parameter is selected as 35 pixels. In the OCR recognition process of traffic electronic police pictures, this parameter is selected as 30 pixels.

2) Select an image containing characters to be recognized in the natural environment, and calculate the image difference boundary map in the form of the mean difference template as shown in Figure 2, and the difference boundary map includes a character area and a background area.

3) Thresholding the differential boundary map of the image to be recognized to form a three-valued boundary map of the image to be recognized, the image to be recognized includes a region to be recognized and a background region, and the difference map of the image to be recognized is according to The following formula is used for ternaryization:

$\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; h</span>}\mathrm{<span\; class="notranslate">\; r</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; the\; s</span>}\mathrm{<span\; class="notranslate">\; h</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; l</span>}\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; \&CenterDot;</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; \&Center\; Dot;</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; P</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; r</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; m</span>}\mathrm{<span\; class="notranslate">\; A</span>}\\ \mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; P</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">\; P</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; r</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; m</span>}\mathrm{<span\; class="notranslate">\; B</span>}\\ \mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; h</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; r</span>}\mathrm{<span\; class="notranslate">\; the\; s</span>}\end{array}\right.$

Among them, P(x, y) is the difference value of the pixel point (x, y) in the difference map, ParamA and ParamB are preset thresholds, and the value range of c1, c2 and c3 is [0, 255], where c1 Indicates the boundary where the brightness of the current pixel is brighter than the surrounding, c2 indicates the boundary where the brightness of the current pixel is darker than the surrounding, c3 indicates the value of the non-boundary. In the project, in order to facilitate debugging, the values of the three constants should have relatively large differences as much as possible. The color of c1 is black, and the value is c1=0; the color of c2 is white, and the value is c2=255, the color of c3 is gray, the value is c3=128, the values of the difference border pixels are 0, 255 respectively, and the value of non-border pixels is 128, that is, the black and white pixels Points are boundary pixels, and the gray pixels are non-boundary pixels.

4) In the three-valued boundary map of the image to be recognized obtained above, search according to the slope angle range and the straight line boundary, and check whether the boundary exceeds the valid range of the stroke length, if it does not exceed the valid range of the stroke length, It is considered to be an effective boundary point and reserved. If it exceeds the effective range of the stroke length, the boundary is judged as a noise boundary, and it is removed from the three-valued boundary image, as shown in Figure 3, which shows three detection processes , starting from the differential boundary pixel points A, C and E respectively, the linear boundaries in Num directions are detected. In Figure 3, there are two linear differential boundaries, namely AB and CD, and E is an isolated , if AB and CD exceed the valid range of the stroke length, their color values are set to the non-boundary value c2, thus clearing the linear boundary.

The difference image processed by the above steps is the image with the adhesion noise eliminated, and it can be applied to the subsequent process of OCR character recognition at this time.

In the above-mentioned process, the search of linear boundary is particularly important, as shown in Figure 4, it is the flow chart that the present invention carries out linear boundary search to three-valued image, and described process is as follows:

Step 4.1) Calculate the linear tilt offset detection template with the tilt angle in the range of [α, β]. The two parameters involved here actually represent the tilt angle range of the OCR character, the minimum tilt angle is α, and the maximum tilt angle is β, their selection has an important impact on the noise removal effect and performance, and it should be set as a configurable parameter in the actual application process. When this method is applied to the OCR recognition of bayonet license plate images, the parameter selection can be α= -15°, β=15°, when applied to the OCR recognition of traffic electronic police pictures, the recommended parameter selection is α=-25°, β=25°, when applied to the OCR recognition of traffic violation parking forensics pictures, due to The inclination angle of illegal parking signs is often relatively large, and the parameters are recommended to be α=-45°, β=45°;

Step 4.2) For each boundary point in the three-valued image, use the straight line detection template described in step 4.1) to check the continuous boundary line starting from the boundary point, having the same boundary value and matching the template;

Step 4.3) For the matched continuous boundary line, check whether its length belongs to the valid range. If it belongs to the valid range, keep it, otherwise the boundary line is a noise boundary, and set the boundary pixel values contained in all boundary points on the boundary line to c3, the effective range of the length of the continuous boundary line mentioned here refers to the stroke The valid range of lengths.

For the aforementioned straight line detection template, its purpose is to speed up the search, so that the entire algorithm meets the real-time requirements in practical applications. There are several straight line detection templates, and each template corresponds to a specific inclination angle. The information is the offset between each y coordinate on the inclined line and the horizontal line. Figure 5 is a flow chart of the line detection algorithm. The specific operations are as follows:

Step 4.1.1) According to the longest stroke length StrokeLen set in the step 1), calculate the number of straight line templates Num, the length Len of the straight line template lines and the maximum offset value MaxOffY of y according to the actual situation of the project. Calculated as follows:

$\left\{\begin{array}{c}\mathrm{<span\; class="notranslate">\; N</span>}\mathrm{<span\; class="notranslate">\; u</span>}\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 4</span>}\mathrm{<span\; class="notranslate">\; S</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; r</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; k</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; no</span>}\\ \mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; S</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; r</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; k</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; no</span>}\\ \mathrm{<span\; class="notranslate">\; m</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; x</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; no</span>}\end{array}\right.<span\; class="notranslate">\; ;</span>$

Step 4.1.2) For each straight-line tilt offset template, calculate the offset value in the y direction according to the following formula.

$\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; *</span>\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; Y</span>}}{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; \&Element;</span><span\; class="notranslate">\; \&lsqb;</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; \&rsqb;</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; \&Element;</span><span\; class="notranslate">\; \&lsqb;</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; m</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; x</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; m</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; x</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; \&rsqb;</span>$

The search for the linear boundary in the three-valued image is the key of this method. For each differential boundary point, starting from this point, use the straight line detection template to match and check whether there is a generalized linear boundary composed of continuous differential boundary pixels. For generalized straight line detection, the method is as follows:

4.2.1) When the straight line is an ordinary straight line with a line width of 1, the method starts from the starting point and checks from left to right whether there is a directly adjacent equivalent boundary point. If there is, the boundary point belongs to the matching point. Otherwise, it is a non-matching point. As shown in Figure 6, under the definition of ordinary straight lines, there are two boundary straight lines in this figure, namely AB and CD.

4.2.2) When the line width of the straight line is N, based on an eight-connected generalized straight line (as shown in Figure 7), the method starts from the starting point and detects the equivalent difference boundary points in the eight-connected range, if there is such boundary point, the boundary point is a matching point, otherwise it is a non-matching point. The meaning of the eight-connectedness described here is shown in Figure 7. If two differential boundary points satisfy the positional relationship shown therein, then these two differential boundary points are eight-connected adjacent. According to this definition, as shown in Figure 8 , 9, which gives examples of two generalized straight lines. In Figure 8, the generalized straight line AB extends from the starting point A to B, with a width of 2 and a horizontal length of 13. In Figure 9, the generalized straight line CD Extending from the starting point C to D, the width is 3, and the horizontal length is 16. The generalized straight line defined in this way has a high degree of matching for the adhesion noise in the project. Clear such a generalized straight line from the differential boundary map It can greatly reduce the influence of noise, especially the adhesion noise, and improve the accuracy of OCR character positioning. In practical engineering applications, the parameter N is usually less than 4.

The embodiment of the present invention proposes a method for eliminating the adhesion noise between characters and the background of the image OCR recognition system. According to the actual situation of the project, after setting the corresponding parameters, then use the mean value difference template to calculate and obtain the three-valued difference map of the image, and finally, Search for a generalized straight line that satisfies the relevant conditions in all differential boundaries. This straight line is the interference noise. By eliminating the differential boundary points related to the generalized straight line, the noise information in the picture to be recognized is greatly suppressed, and finally the candidate character area is accurately identified. Positioning, thereby improving the final correct rate of OCR character recognition.

As shown in Figure 10-19, it is the effect diagram of the application of the present invention in license plate recognition, in which 5 license plate images with adhesion are given, as shown in Figure 10, the four characters in the middle of the license plate are connected due to the application of adhesion noise Together, if the license plate location is performed according to the traditional connected components, the connected components formed by these four characters will be considered as a whole, and after the application of the present invention, as shown in Figure 11, these four characters are completely separated , so as to facilitate the positioning of the license plate, as described in Figures 12, 14, 16, and 18, these license plates also have similar effects, that is, after being processed by the present invention, as shown in Figures 13, 15, 17, and 19, they can all be effective Eliminate the cohesive noise so that the characters in it can be clearly segmented and processed, thereby improving the recognition rate of the system.

The character and background cohesion noise elimination method of a kind of picture OCR recognition system of this embodiment, aim at the OCR technology of existing natural scene picture, the deficiency that exists in the aspect of noise processing, propose a kind of new cohesion noise elimination method, in complex In the changeable natural environment, this method solves the problem that the existing OCR application system has difficulty in locating characters with cohesive noise, realizes accurate extraction of character positions, and can meet the real-time requirements of engineering applications.

Those of ordinary skill in the field of the present invention can understand that the above embodiment of the present invention is only one of the preferred embodiments of the present invention, and is limited by space. All implementation modes cannot be listed here one by one, and any implementation that can embody the technical solution of the claims of the present invention , are all within the protection scope of the present invention.

It should be noted that the above content is a further detailed description of the present invention in conjunction with specific embodiments, and it cannot be determined that the specific embodiments of the present invention are limited thereto. Under the guidance of the present invention, those skilled in the art can Various improvements and modifications are made on the basis of the invention, and these improvements or modifications fall within the protection scope of the present invention.

Claims (6)

1. the character of picture OCR recognition system and a background adhesion noise cancellation method, is characterized in that: character and the background adhesion noise cancellation method of described picture OCR recognition system comprise as follows:

1) priori of the engineering-environment applied according to OCR system, arranges the effective range of the stroke length StrokeLen of character to be identified;

2) choose the image containing character to be identified under physical environment, and with the mode computed image difference boundary graph of mean difference template, described difference boundary graph comprises character zone and background area;

3) the difference boundary graph of image to be identified obtained above is carried out thresholding, form the three-valued boundary graph of image to be identified, described three-valued boundary graph comprises region to be identified and background area;

4) in the three-valued boundary graph of image to be identified obtained above, according in range of tilt angles, linear boundary searches for, check this border whether beyond the effective range of stroke length, when not exceeding the effective range of stroke length, then think effective frontier point, and retain, when exceeding the effective range of stroke length, then this edge determination is noise, it is removed from three value boundary images.

2. the character of a kind of picture OCR recognition system according to claim 1 and background adhesion noise cancellation method, is characterized in that: described step 3) in, by as follows for the formula that the difference boundary graph of described image to be identified carries out thresholding:

$Threshold\left(P\right(x,y\left)\right)=\left\{\begin{array}{cc}c1,& ifP(x,y)<ParamA\\ c2,& ifP(x,y)>ParamB\\ c3,& Others\end{array}\right.$

Wherein, P (x, y) is the difference value of pixel (x, y) in difference diagram, ParamA and ParamB is the threshold value pre-set, c1, c2 and c3 are signless integers, and span is [0,255], wherein c1 represents border brighter around the brightness ratio of current pixel point, and c2 represents the border that the brightness ratio of current pixel point is darker comparatively around, and c3 represents the value on non-border.

3. the character of a kind of picture OCR recognition system according to claim 2 and background adhesion noise cancellation method, it is characterized in that: described c1 color is black, value is c1=0; Described c2 color is white, value is c2=255, described c3 color is grey, value is c3=128, the value of difference boundary pixel point is respectively 0,255, but not the value of boundary pixel point is 128, the pixel of namely described black and white is boundary pixel point, and the pixel of described grey is non-boundary pixel point.

4. the character of a kind of picture OCR recognition system according to claim 1 and background adhesion noise cancellation method, it is characterized in that: described step 4) in, by the three-valued image of described image to be identified according in range of tilt angles, the linear boundary algorithm steps that carries out searching for is as follows:

4.1) calculate the straight line declining displacement of angle of inclination in [α, β] scope and detect masterplate;

4.2) to each frontier point in three-valued image, the straight-line detection masterplate obtained in utilizing step a), checks from this frontier point, boundary value identical, and with the continuum boundary line of stencil matching;

4.3) to the continuum boundary line of coupling, checking whether its length belongs to effective range, when belonging to effective range, then retaining, otherwise this boundary line is noise margin, and the border pixel values that frontier points all on boundary line comprise is set to c3.

5. the character of a kind of picture OCR recognition system according to claim 3 and background adhesion noise cancellation method, it is characterized in that: described step 4-1) in, the algorithm steps of the straight-line detection masterplate of described calculating angle of inclination in [α, β] scope is as follows:

4.1.1) according to described step 1) in set the longest stroke length StrokeLen, according to engineering practice calculated line masterplate quantity Num, the length Len of straight line masterplate lines and the maximum deviation value MaxOffY of y.Computing formula is as follows:

$\left\{\begin{array}{c}Num=4StrokeLen\\ Len=2StrokeLen\\ MaxOffY=Len\end{array}\right.;$

4.1.2) to each straight line declining displacement masterplate, the off-set value in y direction is calculated as follows:

$y=\frac{x*offY}{Len},x\&lsqb;0,Len\&rsqb;,OffY\&Element;\&lsqb;-MaxOffY,MaxOffY\&rsqb;.$

6. the character of a kind of picture OCR recognition system according to claim 3 and background adhesion noise cancellation method, it is characterized in that: described step 4.2) in the three-valued image of image to be identified carried out the searching algorithm of linear boundary, judge from certain frontier point, boundary value identical, and as follows with the method for the continuum boundary line of straight line declining displacement stencil matching:

4.2.1) when the live width of described straight line is the generic linear of 1, the method, from starting point, from left to right checks whether the equivalent boundary of direct neighbor, works as existence, then on this straight line, all frontier points all belong to match point, otherwise just belong to not match point;

4.2.2) when described straight line live width is N, based on the general line of eight connectivity, the method is from starting point, detect the equivalent boundary under eight connectivity meaning, when there is such general line, then on this general line, all frontier points all belong to match point, otherwise just belong to not match point.