CN108038481A - A kind of combination maximum extreme value stability region and the text positioning method of stroke width change - Google Patents

Abstract

The invention relates to a text method combining the maximum extreme value stable area and stroke width variation, using MSER to detect text on images; then performing edge processing on images; and calculating stroke width values along the gradient direction of edge pixels; using form Remove the noise, fill the gap and calculate the connected domain; finally filter the non-text domain and merge the connected domain according to the rules. The advantage of the present invention is that: the present invention obtains the rough text domain through MSER detection, and combines the edge processing, the stroke width change feature and the morphological operation to realize the text positioning in the natural scene image. It is proved by experiments that the present invention has a high accuracy rate , which is beneficial to the subsequent text segmentation and text recognition work, has very obvious practical significance in the field of natural scene text positioning, and can be widely used.

Description

A text localization method combining maximum extremum stable region and stroke width variation

technical field

The invention relates to image processing in the field of artificial intelligence computers, in particular to a method for realizing text positioning in natural scenes by using image processing.

Background technique

In the process of text positioning in natural scenes, there is a basic and unavoidable problem: for images with complex natural backgrounds, how to avoid the influence of factors such as text layout, font type, light intensity, and shooting angle, and accurately obtain text Location.

Text positioning is crucial in the text detection process, and the quality of the text positioning effect directly determines the accuracy of subsequent text segmentation and text recognition. Text localization is more and more widely used in natural scenes, but the complex natural scene environment brings many challenges to this technique. Different from the traditional text positioning technology, there are a lot of distracting objects in the natural scene. At the same time, factors such as shooting angle and font will deform the text, making text positioning more difficult. Therefore, it is necessary to find text features so that the text localization process is not affected by these factors.

There are many methods for text localization in natural scenes, mainly divided into two categories: sliding window method and connected domain analysis method. The sliding window method uses a moving window to detect text in all positions of the image, and the connected domain analysis method obtains candidate connected domains by selecting image features, and realizes text positioning after screening and merging.

In natural scene text positioning, several problems are often encountered that need to be solved:

1) Text feature extraction is a step of text positioning in a natural scene. Therefore, image preprocessing is required before extracting text features, and then the required text features in the image are extracted, and candidate connected domains are generated based on the extracted features.

2) How to distinguish between text domains and non-text domains. There are a large number of distracting objects with very similar characteristics to texts in natural scenes, such as plants, road signs, railings, etc. Therefore, after obtaining the candidate connected domains, it is necessary to distinguish the non-text domains and filter them out.

3) Text forms in natural scenes are diverse, including different fonts and languages. Therefore, how to make the positioning method compatible with various languages and fonts is a problem to be solved.

Therefore, in order to achieve text positioning in natural scenes and obtain a high accuracy rate, the following problems need to be considered:

Technical problem 1: Extraction of text features after natural scene image preprocessing. How to select the text features to be extracted, so that the positioning method can effectively overcome the interference factors of natural scenes and the compatibility of multiple fonts;

Technical problem 2: Candidate connected domain filtering problem. How to design rules to generate candidate connected domains and filter to distinguish non-text domains;

Technical problem 3: The problem of merging single-word connected domains. How to filter single-word connected domains and merge them into text domains;

Aiming at the above difficulties and problems, the present invention proposes and implements a natural scene text positioning method that combines the maximum extreme value stable region and stroke width variation features.

Contents of the invention

The technical problem to be solved by the present invention is to provide a text positioning method that combines the maximum extreme value stable area and stroke width variation, so as to realize accurate text positioning in natural scene images.

In order to solve the above-mentioned technical problems, the technical solution of the present invention is: a text positioning method combining the maximum extreme value stable area and stroke width variation, and its innovative point is that the text positioning method includes the following steps:

(1) Use MSER to detect the text field: grayscale the original image, and use an integer of 0-255 to represent the grayscale value of each pixel in the image; choose a threshold within the grayscale value range of the image, Pixels whose gray value is less than the threshold are defined as black, and pixels greater than the threshold are white. When the threshold is 0, the entire image is white. When the threshold changes from 0 to 255, the black area is stable and the area gradient is the smallest. , then this area is the maximum stable extremum area;

(2) Canny operator marginalizes the image: use the Gaussian filter to smooth the image, calculate the gradient magnitude and gradient direction of the filtered image, suppress the non-maximum value of the gradient magnitude, and find the local maximum in the image gradient Value points, and set the non-local maximum points to zero, so as to refine the image edge, and use the double threshold algorithm to detect and connect the edges;

(3) Obtain image stroke width features: for each edge pixel point, define a ray in the gradient direction perpendicular to the edge, find another corresponding edge pixel point along the ray direction, and find another edge pixel point in the gradient direction , and the gradient direction of this point is approximately opposite to the original gradient direction, then the distance between the two edge pixels is considered as the stroke width; if no corresponding pixel is found or the gradient direction of the corresponding pixel is not approximately opposite, the ray is discarded, and in In a more complex stroke environment, calculate the median stroke width m of all pixels along the undiscarded ray, and set the median stroke width of all pixels on the ray whose stroke width is greater than m to m;

(4) Morphological operation processing image: use opening and closing operations on the image. The opening operation first corrodes the image to remove the edge burrs of the image, and then performs the expansion operation to fill the small gaps and holes in the image. The closing operation first performs expansion Operation, fill the broken area of the image, outline the gap, and then perform an erosion operation to smooth the edge of the image;

(5) Candidate text domain generation: the text pixels are gathered into candidate text domains according to the rules, and the stroke width values of adjacent pixels within the threshold range are classified as the same connected domain, and the aspect ratio and area ratio of the connected domain are calculated, and the Connected domains exceeding the threshold range are filtered out as non-text domains;

(6) Text field merging: the single-character text field is further filtered, and the stroke width average ratio, height ratio, and pixel point color average ratio in the adjacent single-character text field exceed the threshold value, and the connected domain with a large deviation is filtered out as noise, and The remaining connected domains are aggregated into links to form continuous text domains.

Further, in the step of using MSER to detect the text field, the maximum stable extremum region is an algorithm that relies on the relationship between the interior of the region and the boundary pixels, and obtains the maximum stable extremum region according to the stability determination conditions; for The input image is grayscaled, and a threshold value is randomly selected within the gray value range of 0-255. Q1,..., Qi,... are a series of nested extreme value regions, and If q(i)=|Q _i+Δ \Q _i-Δ |/|Q _i | has a local minimum at i ^* , then Q _i* is the maximum extremum stable region MSER.

Further, the Canny operator edge detection is an edge detection operator based on the idea of optimization. This algorithm uses a suitable two-dimensional Gaussian function to smooth and denoise the image by row and column respectively, and calculates the magnitude and direction of the image gradient. , and by suppressing the maximum value of the gradient amplitude, find the local maximum point in the image gradient, set the non-local maximum point to zero, so that the edge can be refined, use T ₁ , T ₂ double threshold algorithm for detection, and use T ₁ is used to obtain each line segment, and T ₂ is used to find breaks on both sides of the line segment and connect the edges; wherein, the two-dimensional Gaussian function is:

I(x,y)=G(x,y)*f(x,y);

The calculation formulas of the gradient magnitude and gradient direction are:

θ(x, y)=arctan(g _y /g _x ); where σ is the standard deviation of the Gaussian curve, and (g _x , g _y ) represents the gradient.

Further, in the step of calculating the stroke width, the value of the stroke width is d _swt ; the step of calculating the value of the stroke width includes: calling the gradient direction of each edge pixel point p d _p , and the gradient direction d _p Perpendicular to the edge direction, define a ray r=p+n d _p , n>0, find another edge pixel point q along the ray direction, if the gradient direction d _q of q is approximately opposite to d _p (d _q =-d _q +π/6), then the pixel stroke width value d _swt is: Among them, x _p and y _p are the horizontal and vertical coordinates of the pixel point p respectively, and x _q and y _q are the horizontal and vertical coordinates of the pixel point q respectively; in a more complex stroke environment, the stroke width value obtained by the above calculation process It is not accurate. Calculate the median stroke width m of all pixels along the undiscarded ray, and set the median stroke width of all pixels on the ray whose stroke width is greater than m to m.

Further, the step of using morphological operations to process the image mainly includes opening and closing operations. The opening operation first performs an erosion operation on the image to remove edge burrs of the image, and then performs an expansion operation to fill small gaps and small holes in the image , the closing operation first performs the expansion operation to fill the fractured area of the image and the contour gap, and then performs the erosion operation to smooth the edge of the image; the opening operation is denoted as defined as: The closed operation is denoted as A·B, which is defined as: Where A is an image and B is a structural element.

Further, in the step of generating the candidate text domains, the non-text domains are mainly filtered out by calculating the attributes of the connected domains and setting rules and thresholds, the rules included include: stroke width variance, aspect ratio, area ratio; the stroke Width variance is used to judge whether the pixel points belong to the same connected domain. If the stroke width values are similar, the pixel is classified as the same connected domain. The calculation formula of stroke width value mean μ _swt and variance σ _swt ² is: where N is the total number of pixels in the connected domain, is the stroke width value of the i-th pixel; the aspect ratio is used to filter out the small and narrow connected domains caused by noise interference, the connected domain aspect ratio r=d _height /d _width , and the aspect ratio threshold is 2. The area ratio is used to filter the connected domains whose area is too large or too small, and the area ratio threshold of the connected domains is 2.

Further, in the step of merging the text domains, the individual character candidate domains are further screened, and the remaining individual character connected domains are gathered into chains to form continuous text domains. The screening conditions of the single character connected domains include stroke width ratio, height ratio, Color mean value difference; The stroke width ratio is used to judge whether adjacent single-character text domains belong to the same text domain, and the stroke width ratio threshold of adjacent single-character text domains is 2; The height ratio is used to judge whether adjacent single-character text domains belong to the same text domain For text fields in the horizontal direction, the height ratio threshold of adjacent single-character text fields is 2; the color mean value is used to determine whether adjacent single-character text fields belong to the same text field, and the color mean difference threshold of adjacent single-character text fields is 40.

The advantages of the present invention are: the present invention utilizes the affine invariance of the maximum extremum stable area to perform MSER text detection on images to obtain multiple candidate text domains; on this basis, the image is edged through the Canny operator; for all edges Pixel points extract stroke width features to obtain connected domains; then further filter non-text domains, merge single-character text domains, and realize text positioning in natural scenes. Experiments have proved that the invention has high accuracy and can be widely used. For example, after combining text segmentation and text recognition, the purpose of text detection in natural scenes can be well realized, and it has very obvious practical significance in the field of image processing.

The present invention uses the test data of the ICDAR2003 text positioning competition data set for experiments, and the experimental results show that: the method provided by the present invention combined with the maximum extreme value stable region and stroke width features can effectively perform text positioning in natural scenes. After statistical analysis, it is known that the natural scene text positioning method combined with the maximum extremum stable region and stroke width features provided by the present invention has a positioning accuracy rate of 74.1%.

Description of drawings

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

Fig. 1 is a flow chart of the text positioning method combined with the maximum extreme value stable area and stroke width variation according to the present invention.

Detailed ways

The following examples can enable those skilled in the art to understand the present invention more comprehensively, but the present invention is not limited to the scope of the described examples.

Example

As shown in Figure 1, the text positioning method combined with the maximum extremum stable region and the stroke width feature provided by the present embodiment includes the following steps:

1. The steps of using MSER to perform text detection on the input image, including:

The extraction of the stroke width feature depends on the edge feature effect of the image. The invention combines MSER to detect the text of the image to obtain a rough text position, which improves the accuracy of image edge and stroke width feature extraction.

The maximum stable extremum region is an algorithm that depends on the relationship between the interior of the region and the boundary pixels. For the grayscale image I: Its maximum stable extremum region is defined as: S is a total order, S=(0,1,...,255}, and satisfies antisymmetry, transitivity, and completeness; define the neighborhood relationship of 4 neighborhoods When p, q∈D, if satisfy p, q are adjacent, denoted as pAq; region Q is a continuous subset of D, for any p, q∈Q, there is a sequence p, a ₁ , a ₂ ,..., a _n , q, such that pAa ₁ ,...,a _i Aa _i+1 ,...,a _n Aq; the boundary of the region The edge of region Q is a set of pixels, at least one pixel is adjacent to the area Q but does not belong to the area Q; the extremum area is a region, for all p ∈ Q, If I(p)>I(q), the area Q is a maximal area, otherwise it is a minimal area; let Q ₁ ,...,Q _i-1 ,Q _i ,... be a series of nested extremum areas, we have If q(i)=|Q _i+Δ \Q _i-Δ |/|Q _i | has a local minimum at i ^* , then Q _i* is the maximum extremum stable region MSER.

2. The steps of Canny operator edge processing image, including:

Use a suitable two-dimensional Gaussian function to smooth and denoise the image by row and column respectively, and calculate the magnitude and direction of the image gradient;

The two-dimensional Gaussian function is:

I(x,y)=G(x,y)*f(x,y) (2);

The calculation formulas of the gradient magnitude and gradient direction are:

θ(x, y) = arctan(g _y /g _x ) (4);

Where σ is the standard deviation of the Gaussian curve, (g _x , g _y ) represents the gradient;

Perform non-maximum value suppression processing on the gradient image, and compare the gradient magnitude of the 8 neighborhoods of each pixel along the gradient direction. If the amplitude values of two pixels in the gradient direction are both smaller than the amplitude value of the pixel itself, then the pixel may be an edge pixel, otherwise the gradient amplitude of the pixel is set to 0. The low threshold t ₁ and high threshold t ₂ are calculated according to the gradient histogram, and the image is thresholded twice according to t ₁ and t _2. If the gradient is smaller than the threshold, its gray value is set to 0.

3. The extraction steps of the stroke width feature, including:

Set the initial stroke width value of each element to infinity. After the edge information is obtained by the Canny operator, the gradient direction of each edge pixel point p is called _dp . Since the edge pixel point p is on the edge, Therefore, the gradient direction d _p must be perpendicular to the edge direction; define a ray r=p+n dp, n>0, find another edge pixel point q along this ray direction, if the gradient direction d _q of q is the same as d _p The directions are approximately opposite (d _q =-d _p +π/6), then the pixel stroke width value d _swt is:

If the corresponding edge pixel point q is not found or the gradient direction d _q and d _p of the edge pixel point q are not opposite, the ray r is discarded;

However, in a more complex stroke environment such as a stroke corner, the stroke width value obtained according to the above calculation process is not accurate, so it is necessary to calculate the median stroke width m of all pixels along all undiscarded rays , set the median stroke width of all pixels larger than m on the ray to m.

4. Steps for image processing by morphological operations, including:

The open operation can make the edge of the image smoother, remove some jagged burrs on the edge, and remove narrow areas. On the contrary, the closing operation operation can remove the noise in the area, fill the narrow and broken part and the gap at the edge. Assuming that there are image A and set B in the integer space Z, the opening operation of B to A is recorded as defined as:

Correspondingly, the closing operation of structure element B on image A is recorded as A·B, which is defined as:

Where A is an image and B is a structural element.

5. The steps of candidate text domain generation, including:

In the step of generating candidate text domains, non-text domains are mainly filtered out by calculating connected domain attributes and setting rules and thresholds. The rules included are: stroke width variance, aspect ratio, and area ratio;

Described stroke width variance is used for judging whether pixel point belongs to same connected domain, if stroke width value is similar, then this pixel point is classified as same connected domain, the calculation formula of stroke width value mean value μ _swt and variance σ _swt ² is:

where N is the total number of pixels in the connected domain, is the stroke width value of the i-th pixel;

The aspect ratio is used to filter out small and narrow connected domains caused by noise interference, the connected domain aspect ratio r=d _height /d _width , and the aspect ratio threshold is 2;

The area ratio is used to filter the connected domains whose area is too large or too small, and the area ratio threshold of the connected domains is 2.

6. Steps for merging text fields, including:

Perform further screening on the single-character candidate domain, and gather the remaining single-character connected domains into a chain to form a continuous text domain. The screening conditions for the single-character connected domain include stroke width ratio, height ratio, and color mean difference;

The stroke width ratio is used to judge whether adjacent single-character text domains belong to the same text domain, and the stroke width ratio threshold value of adjacent single-character text domains is 2;

The height ratio is used to judge whether adjacent single-character text domains belong to the same horizontal direction text domain, and the height ratio threshold of adjacent single-character text domains is 2;

The color mean value is used to judge whether adjacent single-character text fields belong to the same text field, and the color mean value difference threshold of adjacent single-character text fields is 40.

Experiment: Utilize the present invention to adopt ICDAR2003 text localization competition data set test data to carry out experiment. Experimental results show that the method provided by the present invention that combines the features of maximum extreme value stable area and stroke width can effectively locate text in natural scenes. After statistical analysis, it is known that the natural scene text positioning method combined with the maximum extremum stable region and stroke width features provided by the present invention has a positioning accuracy rate of 74.1%. It can be seen from the experimental results that the present invention can effectively realize text positioning in natural scenes, has high accuracy and has very wide application value.

The basic principles and main features of the present invention and the advantages of the present invention have been shown and described above. Those skilled in the industry should understand that the present invention is not limited by the above-mentioned embodiments. What are described in the above-mentioned embodiments and the description only illustrate the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention will also have Variations and improvements are possible, which fall within the scope of the claimed invention. The protection scope of the present invention is defined by the appended claims and their equivalents.

Claims (7)

1. a text positioning method in conjunction with maximum extremum stable region and stroke width variation, is characterized in that: described text positioning method comprises the steps: (1) Use MSER to detect the text field: grayscale the original image, and use an integer of 0-255 to represent the grayscale value of each pixel in the image; choose a threshold within the grayscale value range of the image, Pixels whose gray value is less than the threshold are defined as black, and pixels greater than the threshold are white. When the threshold is 0, the entire image is white. When the threshold changes from 0 to 255, the black area is stable and the area gradient is the smallest. , then this area is the maximum stable extremum area; (2) Canny operator marginalizes the image: use the Gaussian filter to smooth the image, calculate the gradient magnitude and gradient direction of the filtered image, suppress the non-maximum value of the gradient magnitude, and find the local maximum in the image gradient Value points, and set the non-local maximum points to zero, so as to refine the image edge, and use the double threshold algorithm to detect and connect the edges; (3) Obtain image stroke width features: for each edge pixel point, define a ray in the gradient direction perpendicular to the edge, find another corresponding edge pixel point along the ray direction, and find another edge pixel point in the gradient direction , and the gradient direction of this point is approximately opposite to the original gradient direction, then the distance between the two edge pixels is considered as the stroke width; if no corresponding pixel is found or the gradient direction of the corresponding pixel is not approximately opposite, the ray is discarded, and in In a more complex stroke environment, calculate the median stroke width m of all pixels along the undiscarded ray, and set the median stroke width of all pixels on the ray whose stroke width is greater than m to m; (4) Morphological operation processing image: use opening and closing operations on the image. The opening operation first corrodes the image to remove the edge burrs of the image, and then performs the expansion operation to fill the small gaps and holes in the image. The closing operation first performs expansion Operation, fill the broken area of the image, outline the gap, and then perform an erosion operation to smooth the edge of the image; (5) Candidate text domain generation: the text pixels are gathered into candidate text domains according to the rules, and the stroke width values of adjacent pixels within the threshold range are classified as the same connected domain, and the aspect ratio and area ratio of the connected domain are calculated, and the Connected domains exceeding the threshold range are filtered out as non-text domains; (6) Text field merging: the single-character text field is further filtered, and the stroke width average ratio, height ratio, and pixel point color average ratio in the adjacent single-character text field exceed the threshold value, and the connected domain with a large deviation is filtered out as noise, and The remaining connected domains are aggregated into links to form continuous text domains. 2. the text localization method in conjunction with maximum extremum stable region and stroke width variation according to claim 1, is characterized in that: described utilize MSER to carry out in the step that text domain is detected, the maximum stable extremum region is a kind of Relying on the relationship between the interior of the region and the boundary pixels, the algorithm to obtain the maximum stable extreme value region according to the stability determination conditions; for the grayscale of the input image, a threshold is randomly selected within the gray value range of 0-255, Q1, ...., Qi, ... is a series of nested extremum regions, and If q(i)=|Q _i+Δ \Q _i-Δ |/|Q _i | has a local minimum at i ^* , then Q _i* is the maximum extremum stable region MSER. 3. the text localization method that combines maximum extremum stable region and stroke width variation according to claim 1, is characterized in that: described Canny operator edge detection is the edge detection operator based on optimization idea, and this algorithm adopts A suitable two-dimensional Gaussian function smoothes and denoises the image by row and column respectively, calculates the magnitude and direction of the image gradient, and finds the local maximum point in the image gradient by suppressing the maximum value of the gradient magnitude, and sets it to zero The non-local maximum point makes the edge thinner, using the T ₁ and T ₂ double-threshold algorithm for detection, using T ₁ to obtain each line segment, and using T ₂ to find the breaks on both sides of the line segment and connect the edges; among them, The two-dimensional Gaussian function is: <mrow><mi>G</mi><mrow><mo>(</mo><mi>x</mi><mo>,</mo><mi>y</mi><mo>)</mo></mrow><mo>=</mo><mfrac><mn>1</mn><mrow><mn>2</mn><msup><mi>&amp;pi;&amp;sigma;</mi><mn>2</mn></msup></mrow></mfrac><mi>exp</mi><mrow><mo>(</mo><mo>-</mo><mfrac><mrow><msup><mi>x</mi><mn>2</mn></msup><mo>+</mo><msup><mi>y</mi><mn>2</mn></msup></mrow><mrow><mn>2</mn><msup><mi>&amp;sigma;</mi><mn>2</mn></msup></mrow></mfrac><mo>)</mo></mrow><mo>;</mo></mrow> I(x,y)=G(x,y)*f(x,y); The calculation formulas of the gradient magnitude and gradient direction are: <mrow><mi>M</mi><mrow><mo>(</mo><mi>x</mi><mo>,</mo><mi>y</mi><mo>)</mo></mrow><mo>=</mo><msqrt><mrow><msubsup><mi>g</mi><mi>x</mi><mn>2</mn></msubsup><mo>+</mo><msubsup><mi>g</mi><mi>y</mi><mn>2</mn></msubsup></mrow></msqrt><mo>;</mo></mrow> θ(x, y)=arctan(g _y /g _x ); where σ is the standard deviation of the Gaussian curve, and (g _x , g _y ) represents the gradient. 4. the text localization method that combines maximum extremum stable area and stroke width variation according to claim 1, is characterized in that: in the step of described stroke width calculation, described stroke width value is _dswt ; Described stroke width The calculation steps of the value include: call the gradient direction of each edge pixel point p d _p , the gradient direction d _p is perpendicular to the edge direction, define a ray r=p+n·d _p , n>0, find along the ray direction Another edge pixel q, if the gradient direction d _q of q is approximately opposite to d _p (d _q =-d _p +π/6), then the stroke width d _swt of this pixel is: Among them, x _p and y _p are the horizontal and vertical coordinates of the pixel point p respectively, and x _q and y _q are the horizontal and vertical coordinates of the pixel point q respectively; in a more complex stroke environment, the stroke width value obtained by the above calculation process It is not accurate. Calculate the median stroke width m of all pixels along the undiscarded ray, and set the median stroke width of all pixels on the ray whose stroke width is greater than m to m. 5. The text positioning method combined with the maximum extremum stable region and stroke width variation according to claim 1, characterized in that: in the step of processing images using morphological operations, it mainly includes opening and closing operations, and opening operations First perform an erosion operation on the image to remove the edge burrs of the image, and then perform an expansion operation to fill the small gaps and holes in the image. The closing operation first performs the expansion operation to fill the broken area of the image and the contour gap, and then performs an erosion operation to smooth the edge of the image ; The described opening operation is denoted as defined as: The closed operation is denoted as A·B, which is defined as: Where A is an image and B is a structural element. 6. the text localization method that combines maximum extremum stable area and stroke width variation according to claim 1, is characterized in that: in the step of described candidate text domain generation, mainly by calculating connected domain attributes and setting rules and thresholds Filter non-text domains, the rules included are: stroke width variance, aspect ratio, area ratio; the stroke width variance is used to judge whether the pixel points belong to the same connected domain, if the stroke width values are similar, then this pixel point is classified into for the same connected domain. The calculation formula of stroke width value mean μ _swt and variance σ _swt ² is: where N is the total number of pixels in the connected domain, is the stroke width value of the i-th pixel; the aspect ratio is used to filter out the small and narrow connected domains caused by noise interference, the connected domain aspect ratio r=d _height /d _width , and the aspect ratio threshold is 2. The area ratio is used to filter the connected domains whose area is too large or too small, and the area ratio threshold of the connected domains is 2. 7. the text localization method that combines maximum extremum stable area and stroke width variation according to claim 1, is characterized in that: in the step of described text domain merging, further screening is carried out to single character candidate domain, and remaining single character is connected Domains are aggregated into chains to form continuous text domains. The screening conditions for single-character connected domains include stroke width ratio, height ratio, and color mean value difference; the stroke width ratio is used to judge whether adjacent single-character text domains belong to the same text domain. Single-character text field stroke width ratio threshold is 2; Said height ratio is used to judge whether adjacent single-character text fields belong to the same horizontal direction text field, and adjacent single-character text field height ratio threshold is 2; Described color mean value is used for judging adjacent Whether the single-character text fields belong to the same text field, the color mean difference threshold of adjacent single-character text fields is 40.