CN108038486A - A kind of character detecting method - Google Patents

Abstract

The invention discloses a character detection method, which comprises: extracting the extremum region of the character picture to be detected, filtering the extremum region to obtain character candidate regions; calculating MSSH features and deep convolution features, and passing through the self-encoding neural network The MSSH feature and the deep convolution feature are fused to obtain the fusion feature; the character area is further screened out from the character candidate area according to the fusion feature; all the character areas are merged to obtain the final text area. The detection method of the invention has strong robustness, high detection efficiency, and can quickly complete the text detection task.

Description

A Text Detection Method

technical field

The invention relates to a character detection method.

Background technique

As one of the most influential inventions of mankind, writing has played an important role in human life. The rich and precise information contained in text is of great significance to the application of natural scene understanding based on visual semantics. Accurate and robust text detection is required for more and more multimedia applications, such as street scene understanding, traffic sign understanding for driverless cars, and semantic-based image retrieval. The basic task of text detection is to determine whether there is text in scene images and videos, and if so, mark its location. In recent years, with the increase in the capability and number of image acquisition devices, the number of images and videos containing scene text has increased dramatically compared to the past. Therefore, text detection in natural scene images and videos has received more and more attention. With the gradual and in-depth research of computer vision related technologies, how to use computer algorithms for scene text detection has become one of the important and active international frontier topics.

Scene text detection and recognition with low quality and complex backgrounds is extremely challenging. Scene text often has the characteristics of low resolution, complex background, arbitrary orientation, perspective deformation, and uneven lighting, while document text has a uniform format and a single background.

Contents of the invention

The present invention overcomes the deficiencies in the prior art, provides a character detection method, and solves the technical problems of low character detection success rate and weak robustness in the prior art.

In order to solve the above-mentioned technical problems, the technical solution adopted in the present invention is: a kind of character detection method, and this method comprises the steps:

Extract the extreme value area of the text image to be detected, filter the extreme value area, and obtain the character candidate area;

Calculate MSSH features and deep convolution features, and fuse MSSH features and deep convolution features through self-encoding neural network to obtain fusion features;

Further filter out the character region from the character candidate region according to the fusion feature;

Merge all character regions to get the final text region.

The specific method of extracting the extreme value area is as follows:

Convert the text image to be detected into a grayscale image I _gray , an R value image I _R , a G value image I _G and a B value image I _B ;

Calculate the extreme value area for I _R , I _G , and I _B respectively, as follows:

The extreme value area A _R of the R value graph I _R is defined as:

Among them, I _R (p) represents the value of pixel point p in the R value map; I _R (q) represents the value of pixel point q in the R value map; θ represents the threshold value of the extreme value region; Represents a set of pixels adjacent to the extreme value area _AR but not belonging to the extreme value area _AR ;

The extreme value region A _G of the G value graph I _G is defined as:

Among them, I _G (p) represents the value of pixel point p in the G-value map; I _G (q) represents the value of pixel point q in the G-value map; θ represents the threshold value of the extreme value region, Represents a set of pixels adjacent to the extremum area _AG but not belonging to the extremum area _AG ;

The extreme value area A _B of the B value graph I _B is defined as:

Among them, I _B (p) represents the value of pixel point p in the B-value map; I _B (q) represents the value of pixel point q in the B-value map; θ represents the threshold value of the extreme value region, Indicates the set of pixels adjacent to the extremum region A _B but not belonging to the extremum region A _B.

The method of obtaining character candidate regions is as follows:

Calculate the area S, perimeter C, Euler number E, and pixel value variance H of each extreme value region, where the pixel value variance H is calculated through the grayscale image I _gray , and its calculation formula is:

Wherein: x represents a pixel point; I _gray (x) represents the grayscale value of pixel point x; a represents the color interval with the largest number of pixels in the extreme value region; b represents the color interval with the largest number of pixels in the extreme value region; n _a represents the number of pixels in the color interval a in the extreme value region; n _b represents the number of pixels in the color interval b in the extreme value region; R _a represents the set of pixels in the color interval a in the extreme value region; R _b represents the extreme value region The set of pixels in the color interval b in the value area; μ _a represents the average value of the pixel values in the color interval a in the extreme value area; μ _b represents the average value of the pixel values in the color interval b in the extreme value area;

Use the area S, perimeter C, Euler number E, and pixel value variance H of each extreme value area to filter out redundant extreme value areas. After filtering out redundant extreme value areas, the rest is the character candidate area. Filter conditions as follows:

Among them, S ₀ represents the threshold value of the area S of the extreme value region; C ₀ represents the threshold value of the circumference of the extreme value region; E ₀ represents the threshold value of the Euler number in the extreme value region; H ₀ represents the threshold value of the variance of the pixel value in the extreme value region.

The specific method of calculating MSSH features is as follows:

Obtain stroke pixel pairs and stroke line segments of the character candidate region;

Calculate the symmetrical feature description value of a certain stroke pixel pair in the gray value and gradient attribute of the character candidate area;

Calculate the symmetrical feature description of all stroke line segments in the character candidate area on the stroke width value, stroke sequence value distribution and low-frequency pattern attributes;

The eigenvalues of different symmetric attributes are connected to form the MSSH feature, and the specific formula is expressed as follows:

F _m (e _i )=[F _j |=V,G _m ,G _o ,Sw,Md,Pa]

Among them: F _m (e _i ) represents the value of the MSSH feature vector; [] represents the vector connection operation; e _i represents the ith character candidate region; F _j represents the feature vector corresponding to the symmetric attribute; j represents the specific type of the symmetric attribute; V represents the gray value; G _m represents the gradient size attribute; G _o represents the gradient direction attribute; Sw represents the stroke width value; Md represents the stroke sequence value distribution; Pa represents the low-frequency pattern attribute.

The specific method of obtaining the stroke pixel pairs and stroke line segments of the character candidate area is as follows:

Use the Canny edge detection operator to output the edge image;

Calculate the gradient direction of a certain pixel point p on the stroke edge image;

Follow the ray r determined by the gradient direction until the ray meets another stroke edge pixel point q;

A stroke pixel pair is defined as {p, q}, and a stroke line segment is defined as the distance of ray r between pixel points p and q.

The specific method of calculating the depth convolution feature is as follows:

Resize the character candidate area to a 64×64 pixel value;

Construct a three-stage convolutional neural network model;

The one-stage construction method is as follows:

In the first stage, two convolutional layers and a maximum pooling layer are used sequentially. Among them, the convolutional layer uses 32 convolution kernels with a size of 3×3, and 1 pixel is the displacement offset, which is convolved with the character candidate area. Product operation, the specific formula is as follows:

Where g(a,b,k) represents the value of the pixel value of row a and column b in the character candidate area after the kth convolution operation; e _i (a+m,b+n) represents the i-th character candidate The pixel value of row (a+m) and column (b+n) in the area; m represents the row offset of the pixel, n represents the column offset of the pixel, and its value set is {-1,0,1} , h _k represents the kth convolution kernel; after each convolution layer operation, the activation value is calculated using a nonlinear activation function, the specific formula is as follows:

f(a,b,k)=max(0,g(a,b,k))

f(a,b,k) represents the activation value of the pixel value of row a and column b in the character candidate area after the kth convolution operation; max() represents the function of taking a large value;

The activation value is then passed to the maximum pooling layer, which uses 2 pixels as the stride and takes the maximum value in the 2×2 spatial neighborhood as the output value;

The structure of the second stage is the same as that of the first stage;

The three-stage sequence uses three convolutional layers, a maximum pooling layer, and a fully connected layer. The fully connected layer connects the output of the maximum pooling layer into a one-dimensional vector as input, and controls the output to 128 dimensions. The formula can be expressed as follows:

F _d ＝W·X+B

Among them: F _d is the generated 128-dimensional deep convolution feature, X is the one-dimensional vector obtained after connecting the output of the maximum pooling layer, W is the weight matrix, and B is the offset vector;

The convolutional neural network model is trained and tested, the values of unknown parameters h _k , W and B are determined through training, and the F _d generated by the test is used as the deep convolution feature of the character candidate area.

The method of obtaining fusion features is as follows:

Use the weight ω _d of the convolutional neural network model after training as the initial fusion weight value of the deep convolution feature F _d ;

For the MSSH feature F _m , the logistic regression model is used to predict its initial fusion weight value ω _m and reduce its feature dimension size. The specific process can be expressed by the following formula:

in, Represents the MSSH feature after dimensionality reduction, e _i represents the i-th character candidate region; the function f _τ () represents the logistic regression model, representing the small data set used to train the initial weight value of the feature;

The specific process of generating the fusion feature F _s can be expressed by the following formula:

Among them, the function f _μ () represents the self-encoder network, and F _d are consistent in dimension.

During fused training, the joint training process of autoencoder networks ends when the validation error rate stops decreasing.

The specific method of merging character areas is as follows:

Assuming that the character area is S, calculate the center point c _i of all s _i ∈ S;

For any two-character area s _i , s _j ∈ S, if the Euler distance between the two center points is less than the threshold F, connect a straight line l _i, between the two center points;

Calculate the angle α between all the straight lines and the horizontal line, and take the mode α _mode of all the included angles; keep the straight lines whose included angles are in the interval [α _mode -π/6,α _mode +π/6], and remove the other straight lines;

The character regions connected by straight lines are merged to obtain the final text region.

Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

1. Use MSSH features and deep convolution features to describe character candidate areas, where MSSH features are based on edge images, and are robust to low resolution, image rotation, affine deformation, and multi-language and multi-font changes; depth convolution The product feature construction process does not require manual intervention, and has a strong ability to describe the appearance attributes of character candidate areas. In the process of low resolution, image rotation, and illumination changes, the overall appearance of the image does not change much, and it also has strong robustness. ;

(2) The self-encoding network used in the present invention can automatically fuse MSSH features and deep convolution features without manual intervention, and the generated fusion features can integrate the advantages of each feature, which is suitable for low resolution, image rotation, affine deformation and complex The background is robust.

(3) The method for character detection in natural scenes involved in the present invention has high efficiency, and the calculation algorithm complexity is not high, and the character detection process can be completed quickly.

Description of drawings

Fig. 1 is a flow chart of the present invention;

Fig. 2 is a calculation flowchart of the depth convolution feature in Fig. 1;

Fig. 3 is a flowchart of feature fusion in Fig. 1;

Figure 4 is a text picture to be detected;

Fig. 5 is the picture of the character candidate region obtained by Fig. 4 through extremum region filtering;

Fig. 6 is the character region obtained by Fig. 5 through feature fusion;

Fig. 7 is the text area obtained from Fig. 6 by merging character areas.

Detailed ways

The present invention provides a text detection method, by extracting and filtering the extreme value area, the character candidate area is obtained, the character area is further screened out from the character candidate area through the fusion of MSSH feature and depth convolution feature, and finally the character area is merged to obtain the text area. The detection method of the invention has strong robustness, high detection efficiency, and can quickly complete the text detection task.

The present invention will be further described below in conjunction with the accompanying drawings. The following examples are only used to illustrate the technical solution of the present invention more clearly, but not to limit the protection scope of the present invention.

As shown in Figure 1, it is a flowchart of the present invention, and the inventive method specifically includes the following steps:

Step 1: Input the text picture to be detected, and extract the extreme value area of the text picture to be detected;

First, convert the input RGB color image into a gray scale image I _gray , a red component image R value image I _R , a green component image G value image I _G and a blue component image B value image I _B ;

Secondly, for I _R , I _G , and I _B , calculate the extreme value regions _AR , A _G , A _B respectively. The extreme value region refers to the region where the pixel value of the outer boundary of the region is strictly greater than the pixel value in the region, and the R value diagram Taking I _R as an example, its extremum region _AR can be defined as:

Among them, I _R (p) and I _R (q) represent the values of pixel points p and q in I _R respectively, and θ represents the threshold value of the extreme value region, Represents a set of pixels adjacent to the extreme value area _AR but not belonging to the extreme value area _AR ;

Then, calculate the area S, perimeter C, Euler number E, and pixel value variance H of each extreme value region, where the pixel value variance H is calculated through the grayscale image I _gray , and its calculation formula is:

Wherein x represents a pixel point, I _gray (x) represents the gray value of pixel point x, a and b are respectively the color interval with the largest number of pixels in the extreme value region and the color interval with the largest number of pixels, n _a and n _b respectively represent the number of pixels in the color interval a and b in the extreme value region, R _a and R _b represent the pixel sets in the color interval a and b in the extreme value region respectively, μ _a and μ _b represent the pixels in the extreme value region Average of pixel values in color intervals a and b.

Step 2: Filter the extreme value area to obtain the character candidate area;

Use the area S, perimeter C, Euler number E, and pixel value variance H of each extremum region to filter out redundant extremum regions, and what remains after filtering out redundant extremum regions is the character candidate region. The filter conditions are as follows:

Among them, S ₀ , C ₀ , E ₀ , and H ₀ are all thresholds obtained through statistics of a large number of character and non-character regions. S ₀ indicates the threshold value of the area S of the extreme value area, and the specific value of S ₀ is within the interval [80,120; C ₀ indicates the threshold value of the circumference of the extreme value area, and the specific value of C ₀ is within the interval [30,50]; E ₀ indicates the extreme value area The threshold of the Euler number, the specific value of E ₀ is in the interval [0,1]; H ₀ represents the threshold of the variance of the pixel value in the extreme value area, and the specific value of H ₀ is in the interval [100,200].

Fig. 4 is an input character picture to be detected. As shown in Fig. 5, it is a picture of the character candidate area obtained by filtering the extremum area from Fig. 4 .

Step 3: Calculate MSSH features and deep convolution features;

The specific method of calculating MSSH features is as follows:

Obtain stroke pixel pairs and stroke line segments of the character candidate region;

Obtain the stroke pixel pairs and stroke line segments of the character candidate area through the SWT algorithm, the steps are as follows:

(1) Use the Canny edge detection operator to output the edge image;

(2) Calculate the gradient direction of a certain pixel p on the stroke edge image;

(3) Follow the ray r determined by the gradient direction until the ray meets another stroke edge pixel point q;

(4) The stroke pixel pair is defined as {p, q}, and the stroke segment is defined as the distance between the ray r between the pixel points p and q.

The steps of the Canny edge detection algorithm are as follows:

(1) Convert the character candidate area into a grayscale image;

(2) Gaussian filtering is carried out to the obtained grayscale image;

(3) Calculate the magnitude and direction of the gradient;

(4) Perform non-maximum suppression on the gradient amplitude;

(5) Detect and connect edges with a double-threshold algorithm.

Calculate the symmetrical feature description value of a certain stroke pixel pair in the gray value and gradient attribute of the character candidate area;

Assuming {p,q} is a certain stroke pixel pair in the character candidate area, the calculation steps of the symmetrical attribute description value based on the stroke pixel pair are as follows:

(1) Calculate the feature value F _j (p,q) ₁ of the stroke pixel pair {p,q} on the gray value and gradient size attribute by the following formula:

F _j (p,q) ₁ ＝f _h (|I _j (p)-I _j (q)|)if∈{V,G _m }

Among them, I _j (p) represents the value of pixel point p on symmetric attribute j, I _j (q) represents the value of pixel point q on symmetric attribute j; j represents the specific type of symmetric attribute; {V, G _m } respectively refer to the gray value and gradient size attributes, and the function f _h () represents the statistical operation of the histogram.

(2) We calculate the feature value F _j (p,q) ₂ of the stroke pixel pair {p,q} on the gradient direction attribute by the following formula:

F _j (p,q) ₂ ＝f _h (cos<I _j (p), _j (q)>)j=G _o

Among them, G _o refers to the gradient direction attribute, cos<> represents the arccosine function, and the function f _h () represents the histogram statistical operation.

Calculate the symmetrical feature description of all stroke line segments in the character candidate area on the stroke width value, stroke sequence value distribution and low-frequency pattern attributes;

Assuming s means that for the stroke line segment set in the character candidate area, the calculation steps of the symmetric attribute description value based on the stroke line segment set are as follows:

(1) Calculate the eigenvalue F _j (s) of the stroke pixel pair on the gradient direction attribute by the following formula:

F _j (s)=f _h (f _ξ (s,j))j∈{Sw,Md,Pa}

Among them, the function f _h () represents the statistical operation of the histogram, {Sw, Md, Pa} represents the same kind of symmetrical attributes, including the stroke width value Sw, the stroke sequence value distribution Md and the low-frequency pattern attribute Pa, where the function f _ξ (s, j ) can be defined as:

Where |||| refers to the Euclidean distance, D _s and M _s respectively represent the variance and average value of the gray value of the pixels contained in the stroke line segment set s, Represents Haar wavelet transform, k represents the number of wavelet transformation layers, n _l represents the highest scale layer, and ω _k is a predefined weight parameter. Wherein, the specific value of n _l is 1. When k=0, the specific value of ω _k is 0.1; when k=1, the specific value of ω _k is 0.3; when k=2, the specific value of ω _k is 0.5.

(2) For stroke width value, stroke sequence value distribution, and symmetric feature description value on low-frequency mode attributes, scale the above attribute values of a certain text candidate area to a scale between 0 and 1.

The eigenvalues of different symmetric attributes are connected to form the MSSH feature, and the specific formula is expressed as follows:

F _m (e _i )=[F _j |=V,G _m ,o,w,Md,a]

Among them: F _m (e _i ) represents the value of the MSSH feature vector; [] represents the vector connection operation; e _i represents the i-th text candidate region; F _j represents the feature vector corresponding to the symmetric attribute; j represents the specific type of the symmetric attribute; V represents the gray value; G _m represents the gradient size attribute; G _o represents the gradient direction attribute; Sw represents the stroke width value; Md represents the stroke sequence value distribution; Pa represents the low-frequency pattern attribute.

As shown in Figure 2, it is a calculation flow chart of deep convolution features. The method of calculating deep convolution features is as follows:

Resize the character candidate area to a 64×64 pixel value;

Construct a three-stage convolutional neural network model;

The one-stage construction method is as follows:

In the first stage, two convolutional layers and a maximum pooling layer are used sequentially. Among them, the convolutional layer uses 32 convolution kernels with a size of 3×3, and 1 pixel is the displacement offset, which is convolved with the character candidate area. Product operation, the specific formula is as follows:

Where g(a,b,k) represents the value of the pixel value of row a and column b in the character candidate area after the kth convolution operation; e _i (a+m,b+n) represents the i-th character candidate The pixel value of row (a+m) and column (b+n) in the area; m represents the row offset of the pixel, n represents the column offset of the pixel, and its value set is {-1,0,1} , h _k represents the kth convolution kernel; after each convolution layer operation, the activation value is calculated using a nonlinear activation function, the specific formula is as follows:

f(a,b,k)=max(0,g(a,b,k))

f(a,b,k) represents the activation value of the pixel value of row a and column b in the character candidate area after the kth convolution operation; max() represents the function of taking a large value;

The activation value is then passed to the maximum pooling layer, which uses 2 pixels as the stride and takes the maximum value in the 2×2 spatial neighborhood as the output value;

The structure of the second stage is the same as that of the first stage;

The three-stage sequence uses three convolutional layers, a maximum pooling layer, and a fully connected layer. The fully connected layer connects the output of the maximum pooling layer into a one-dimensional vector as input, and controls the output to 128 dimensions. The formula can be expressed as follows:

F _d ＝W·X+B

Among them: F _d is the generated 128-dimensional deep convolution feature, X is the one-dimensional vector obtained after connecting the output of the maximum pooling layer, W is the weight matrix, and B is the offset vector;

The use of the model is divided into two processes, namely the training process and the testing process. The training process is used to determine the values of the unknown parameters h _k , W and B, and the testing process is used to generate the deep convolution feature F _d of the character candidate area.

During the training process, each character candidate region used for training is assigned a label. When the label is 0, it means that the character candidate area is not a character area; when the label is 1, it means that the character candidate area is a character area. The depthwise convolutional feature F _d will be connected to the two-dimensional label vector through full connection, and its values are 0 and 1 respectively. During the training process, when the label value predicted by the neural model for the character candidate region no longer changes, the training ends. The obtained h _k , W and b at the end of the training are respectively used as fixed values of h _k , W and b.

During the test, the F _d generated by the first, second and third stages of the convolutional neural network will be used as the deep convolution feature of the character candidate area.

The output of the maximum pooling layer in the third stage is the 128-dimensional deep convolutional feature F _d , which will be connected to the fully connected layer during the training of the convolutional neural network. The fully connected layer will output whether the character candidate region is a text or non-text label.

Step 4: Merge MSSH features and deep convolution features through self-encoding neural network to obtain fusion features;

As shown in Figure 3, it is a flow chart of feature fusion, including the following steps:

First, in the fusion process, the weight ω _d of the trained convolutional neural network model is used as the initial fusion weight value of the deep convolutional feature F _d ;

Then, for the MSSH feature F _m , use the logistic regression model to predict its initial fusion weight value ω _m and reduce its feature dimension size. The specific process can be expressed by the following formula:

Among them, the function f _τ () represents the logistic regression model, Represents the MSSH feature after dimensionality reduction, and represents the small data set used to train the initial weight value of the feature.

Finally, based on the self-encoder network fusion of MSSH features and deep convolution features, the specific process of generating fusion features F _s can be expressed by the following formula:

Among them, the function f _μ () represents the self-encoder network, Represents the MSSH feature after dimensionality reduction, and F _d are consistent in dimension.

During fused training, the joint training process of autoencoder networks ends when the validation error rate stops decreasing.

Step 5: further filter out the character region from the character candidate region according to the fusion feature;

Input the fusion feature of the character candidate region into the pre-trained logistic regression classifier to judge whether the character candidate region is a real character region.

The training steps of the logistic regression classifier are as follows:

(1) Take the ICDAR 2013 scene data set, a general scene text detection data set, and calculate the fusion features of all candidate character regions of the data set according to the above steps, and use it as a training set.

(2) Input the training set into the logistic regression algorithm for binary classification problem training.

As shown in Figure 6, it is the character area obtained from Figure 5 through feature fusion.

Step 6: Merge all character regions to obtain the final text region.

First, for the character area S, calculate the center points c _i of all s _i ∈ S;

Secondly, for any character area s _i , s _j ∈ S, if the Euler distance between the center points ci and c _j is less than the threshold F, connect a straight line l _{i,j between} the center points _ci and c _j ; Preferably, the value of F is 5.

Then, calculate the included angle α between all straight lines l and the horizontal line, and take the mode α _mode of all included angles. Keep the straight lines whose included angles are in the interval [α _mode -π/6,α _mode +π/6], and remove the rest of the straight lines. As shown in FIG. 7, it is the character area obtained by merging the character areas from FIG. 6.

The above is only a preferred embodiment of the present invention, it should be pointed out that for those of ordinary skill in the art, without departing from the technical principle of the present invention, some improvements and modifications can also be made. It should also be regarded as the protection scope of the present invention.

Claims (9)

1. a text detection method, is characterized in that, the method comprises the steps: Extract the extreme value area of the text image to be detected, filter the extreme value area, and obtain the character candidate area; Calculate MSSH features and deep convolution features, and fuse MSSH features and deep convolution features through self-encoding neural network to obtain fusion features; Further filter out the character region from the character candidate region according to the fusion feature; Merge all character regions to get the final text region. 2. text detection method according to claim 1, is characterized in that, the concrete method of extracting extremum region is as follows: Convert the text image to be detected into a grayscale image I _gray , an R value image I _R , a G value image I _G and a B value image I _B ; Calculate the extreme value area for I _R , I _G , and I _B respectively, as follows: The extreme value area A _R of the R value graph I _R is defined as: <mrow><msub><mi>I</mi><mi>R</mi></msub><mrow><mo>(</mo><mi>p</mi><mo>)</mo></mrow><mo>&gt;</mo><mi>&amp;theta;</mi><mo>&amp;GreaterEqual;</mo><msub><mi>I</mi><mi>R</mi></msub><mrow><mo>(</mo><mi>q</mi><mo>)</mo></mrow><mo>,</mo><mo>&amp;ForAll;</mo><mi>p</mi><mo>&amp;Element;</mo><msub><mi>A</mi><mi>R</mi></msub><mo>,</mo><mi>q</mi><mo>&amp;Element;</mo><mo>&amp;part;</mo><msub><mi>A</mo>mi><mi>R</mi></msub></mrow> Among them, I _R (p) represents the value of pixel point p in the R value map; I _R (q) represents the value of pixel point q in the R value map; θ represents the threshold value of the extreme value region; Represents a set of pixels adjacent to the extreme value area _AR but not belonging to the extreme value area _AR ; The extreme value region A _G of the G value graph I _G is defined as: <mrow><msub><mi>I</mi><mi>G</mi></msub><mrow><mo>(</mo><mi>p</mi><mo>)</mo></mrow><mo>&gt;</mo><mi>&amp;theta;</mi><mo>&amp;GreaterEqual;</mo><msub><mi>I</mi><mi>G</mi></msub><mrow><mo>(</mo><mi>q</mi><mo>)</mo></mrow><mo>,</mo><mo>&amp;ForAll;</mo><mi>p</mi><mo>&amp;Element;</mo><msub><mi>A</mi><mi>G</mi></msub><mo>,</mo><mi>q</mi><mo>&amp;Element;</mo><mo>&amp;part;</mo><msub><mi>A</mo>mi><mi>G</mi></msub></mrow> Among them, I _G (p) represents the value of pixel point p in the G-value map; I _G (q) represents the value of pixel point q in the G-value map; θ represents the threshold value of the extreme value region, Represents a set of pixels adjacent to the extremum area _AG but not belonging to the extremum area _AG ; The extreme value area A _B of the B value graph I _B is defined as: <mrow><msub><mi>I</mi><mi>B</mi></msub><mrow><mo>(</mo><mi>p</mi><mo>)</mo></mrow><mo>&gt;</mo><mi>&amp;theta;</mi><mo>&amp;GreaterEqual;</mo><msub><mi>I</mi><mi>B</mi></msub><mrow><mo>(</mo><mi>q</mi><mo>)</mo></mrow><mo>,</mo><mo>&amp;ForAll;</mo><mi>p</mi><mo>&amp;Element;</mo><msub><mi>A</mi><mi>B</mi></msub><mo>,</mo><mi>q</mi><mo>&amp;Element;</mo><mo>&amp;part;</mo><msub><mi>A</mo>mi><mi>B</mi></msub></mrow> Among them, I _B (p) represents the value of pixel point p in the B-value map; I _B (q) represents the value of pixel point q in the B-value map; θ represents the threshold value of the extreme value region, Indicates the set of pixels adjacent to the extremum region A _B but not belonging to the extremum region A _B. 3. text detection method according to claim 2, is characterized in that, the method for obtaining character candidate region is as follows: Calculate the area S, perimeter C, Euler number E, and pixel value variance H of each extreme value region, where the pixel value variance H is calculated through the grayscale image I _gray , and its calculation formula is: <mrow><mi>H</mi><mo>=</mo><mfrac><mrow><msub><mi>n</mi><mi>a</mi></msub><mo>&amp;CenterDot;</mo><msub><mi>&amp;Sigma;</mi><mrow><mi>x</mi><mo>&amp;Element;</mo><msub><mi>R</mi><mi>a</mi></msub></mrow></msub><msup><mrow><mo>(</mo><msub><mi>I</mi><mrow><mi>g</mi><mi>r</mi><mi>a</mi><mi>y</mi></mrow></msub><mo>(</mo><mi>x</mi><mo>)</mo><mo>-</mo><msub><mi>&amp;mu;</mi><mi>a</mi></msub><mo>)</mo></mrow><mn>2</mn></msup><mo>+</mo><msub><mi>n</mi><mi>b</mi></msub><mo>&amp;CenterDot;</mo><msub><mi>&amp;Sigma;</mi><mrow><mi>x</mi><mo>&amp;Element;</mo><msub><mi>R</mi><mi>b</mi></msub></mrow></msub><msup><mrow><mo>(</mo><msub><mi>I</mi><mrow><mi>g</mi><mi>r</mi><mi>a</mi><mi>y</mi></mrow></msub><mo>(</mo><mi>x</mi><mo>)</mo><mo>-</mo><msub><mi>&amp;mu;</mi><mi>b</mi></msub><mo>)</mo></mrow><mn>2</mn></msup></mrow><mrow><msub><mi>n</mi><mi>a</mi></msub><mo>+</mo><msub><mi>n</mi><mi>b</mi></msub></mrow></mfrac></mrow> Wherein: x represents a pixel point; I _gray (x) represents the grayscale value of pixel point x; a represents the color interval with the largest number of pixels in the extreme value region; b represents the color interval with the largest number of pixels in the extreme value region; n _a represents the number of pixels in the color interval a in the extreme value region; n _b represents the number of pixels in the color interval b in the extreme value region; R _a represents the set of pixels in the color interval a in the extreme value region; R _b represents the extreme value region The set of pixels in the color interval b in the value area; μ _a represents the average value of the pixel values in the color interval a in the extreme value area; μ _b represents the average value of the pixel values in the color interval b in the extreme value area; Use the area S, perimeter C, Euler number E, and pixel value variance H of each extreme value area to filter out redundant extreme value areas. After filtering out redundant extreme value areas, the rest is the character candidate area. Filter conditions as follows: <mfenced open = "{" close = ""><mtable><mtr><mtd><mrow><mi>S</mi><mo>&amp;le;</mo><msub><mi>S</mi><mn>0</mn></msub></mrow></mtd></mtr><mtr><mtd><mrow><mi>C</mi><mo>&amp;le;</mo><msub><mi>C</mi><mn>0</mn></msub></mrow></mtd></mtr><mtr><mtd><mrow><mi>E</mi><mo>&amp;le;</mo><msub><mi>E</mi><mn>0</mn></msub></mrow></mtd></mtr><mtr><mtd><mrow><mi>H</mi><mo>&amp;GreaterEqual;</mo><msub><mi>H</mi><mn>0</mn></msub></mrow></mtd></mtr></mtable></mfenced> Among them, S ₀ represents the threshold value of the area S of the extreme value region; C ₀ represents the threshold value of the circumference of the extreme value region; E ₀ represents the threshold value of the Euler number in the extreme value region; H ₀ represents the threshold value of the variance of the pixel value in the extreme value region. 4. text detection method according to claim 1, is characterized in that, the concrete method of calculating MSSH feature is as follows: Obtain stroke pixel pairs and stroke line segments of the character candidate region; Calculate the symmetrical feature description value of a certain stroke pixel pair in the gray value and gradient attribute of the character candidate area; Calculate the symmetrical feature description of all stroke line segments in the character candidate area on the stroke width value, stroke sequence value distribution and low-frequency pattern attributes; The eigenvalues of different symmetric attributes are connected to form the MSSH feature, and the specific formula is expressed as follows: F _m (e _i )=[F _j |j=V, G _m , G _o , Sw, Md, Pa] Among them: F _m (e _i ) represents the value of the MSSH feature vector; [] represents the vector connection operation; e _i represents the ith character candidate region; F _j represents the feature vector corresponding to the symmetric attribute; j represents the specific type of the symmetric attribute; V represents the gray value; G _m represents the gradient size attribute; G _o represents the gradient direction attribute; Sw represents the stroke width value; Md represents the stroke sequence value distribution; hu a represents the low-frequency mode attribute. 5. text detection method according to claim 4, is characterized in that, the specific method that obtains the stroke pixel pair of character candidate area and stroke line segment is as follows: Use the Canny edge detection operator to output the edge image; Calculate the gradient direction of a certain pixel point p on the stroke edge image; Follow the ray r determined by the gradient direction until the ray meets another stroke edge pixel point q; A stroke pixel pair is defined as {p, q}, and a stroke segment is defined as the distance of ray r between pixel points p and q. 6. the text detection method according to claim 4, is characterized in that, the concrete method of calculating depth convolution feature is as follows: Resize the character candidate area to a 64×64 pixel value; Construct a three-stage convolutional neural network model; The one-stage construction method is as follows: In the first stage, two convolutional layers and a maximum pooling layer are used sequentially. Among them, the convolutional layer uses 32 convolution kernels with a size of 3×3, and 1 pixel is the displacement offset, which is convolved with the character candidate area. Product operation, the specific formula is as follows: <mrow><mi>g</mi><mrow><mo>(</mo><mi>a</mi><mo>,</mo><mi>b</mi><mo>,</mo><mi>k</mi><mo>)</mo></mrow><mo>=</mo><munder><mo>&amp;Sigma;</mo><mrow><mi>m</mi><mo>,</mo><mi>n</mi><mo>&amp;Element;</mo><mo>{</mo><mo>-</mo><mn>1</mn><mo>,</mo><mn>0</mn><mo>,</mo><mn>1</mn><mo>}</mo></mrow></munder><msub><mi>e</mi><mi>i</mi></msub><mrow><mo>(</mo><mi>a</mi><mo>+</mo><mi>m</mi><mo>,</mo><mi>b</mi><mo>+</mo><mi>n</mi><mo>)</mo></mrow><msub><mi>h</mi><mi>k</mi></msub><mrow><mo>(</mo><mi>m</mi><mo>,</mo><mi>n</mi><mo>)</mo></mrow></mrow> Among them, g(a, b, k) represents the value of the pixel value of row a and column b in the character candidate area after the kth convolution operation; ei(a+m, b+n) represents the i-th character candidate area In the (a+m) row and the (b+n) column pixel value; m represents the row offset of the pixel, n represents the column offset of the pixel, and its value set is {-1, 0, 1}, h _k represents the kth convolution kernel; after each convolution layer operation, the activation value is calculated using a nonlinear activation function, the specific formula is as follows: f(a,b,k)=max(0,g(a,b,k)) f(a, b, k) represents the activation value of the pixel value of row a and column b in the character candidate area after the k convolution operation; max() represents a large-value function; The activation value is then passed to the maximum pooling layer, which uses 2 pixels as the stride and takes the maximum value in the 2×2 spatial neighborhood as the output value; The structure of the second stage is the same as that of the first stage; The three-stage sequence uses three convolutional layers, a maximum pooling layer, and a fully connected layer. The fully connected layer connects the output of the maximum pooling layer into a one-dimensional vector as input, and controls the output to 128 dimensions. The formula can be expressed as follows: F _d ＝W·X+B Among them: F _d is the generated 128-dimensional deep convolution feature, X is the one-dimensional vector obtained after connecting the output of the maximum pooling layer, W is the weight matrix, and B is the offset vector; The convolutional neural network model is trained and tested, the values of unknown parameters h _k , W and B are determined through training, and the F _d generated by the test is used as the deep convolution feature of the character candidate area. 7. The text detection method according to claim 6, wherein the method for obtaining fusion features is as follows: Use the weight ω _d of the convolutional neural network model after training as the initial fusion weight value of the deep convolution feature F _d ; For the MSSH feature F _m , the logistic regression model is used to predict its initial fusion weight value ω _m and reduce its feature dimension size. The specific process can be expressed by the following formula: <mrow><mo>{</mo><mover><msub><mi>F</mi><mi>m</mi></msub><mo>~</mo></mover><mrow><mo>(</mo><msub><mi>e</mi><mi>i</mi></msub><mo>)</mo></mrow><mo>,</mo><msub><mi>&amp;omega;</mi><mi>m</mi></msub><mo>}</mo><mo>=</mo><msub><mi>f</mi><mi>&amp;tau;</mi></msub><mrow><mo>(</mo><msub><mi>F</mi><mi>m</mi></msub><mo>;</mo><mi>D</mi><mo>)</mo></mrow></mrow> in, Represents the MSSH feature after dimensionality reduction, e _i represents the i-th character candidate region; the function f _τ () represents the logistic regression model, and D represents the small data set used to train the initial weight value of the feature; The specific process of generating the fusion feature _FS can be expressed by the following formula: <mrow><msub><mi>F</mi><mi>s</mi></msub><mrow><mo>(</mo><msub><mi>e</mi><mi>i</mi></msub><mo>)</mo></mrow><mo>=</mo><msub><mi>f</mi><mi>&amp;mu;</mi></msub><mrow><mo>(</mo><msub><mi>&amp;omega;</mi><mi>d</mi></msub><mo>,</mo><msub><mi>F</mi><mi>d</mi></msub><mo>(</mo><msub><mi>e</mi><mi>i</mi></msub><mo>)</mo><mo>,</mo><msub><mi>&amp;omega;</mi><mi>m</mi></msub><mo>,</mo><mover><msub><mi>F</mi><mi>m</mi></msub><mo>~</mo></mover><mo>(</mo><msub><mi>e</mi><mi>i</mi></msub><mo>)</mo><mo>)</mo></mrow></mrow> Among them, the function f _μ () represents the self-encoder network, and F _d are consistent in dimension. 8. The text detection method according to claim 7, wherein during the fusion training process, when the verification error rate stops decreasing, the joint training process of the autoencoder network ends. 9. text detection method according to claim 1, is characterized in that, the concrete method of merging character area is as follows: Assuming that the character area is S, calculate the center point c _i of all s _i ∈ S; For any two-character area s _i , s _j ∈ S, if the Euler distance between the two center points is less than the threshold F, connect a straight line l _{i, i} between the two center points; Calculate the angle α between all straight lines and the horizontal line, and take the mode α _mode of all the included angles; keep the straight lines with included angles in the interval [α _mode -π/6, α _mode +π/6], and remove the rest of the straight lines; The character regions connected by straight lines are merged to obtain the final text region.