WO2021027070A1

WO2021027070A1 - Deep learning-based ellipse recognition method

Info

Publication number: WO2021027070A1
Application number: PCT/CN2019/113241
Authority: WO
Inventors: 徐静; 陈恳; 刘炽成
Original assignee: 清华大学
Priority date: 2019-08-09
Filing date: 2019-10-25
Publication date: 2021-02-18
Also published as: CN110598692A; CN110598692B

Abstract

A deep learning-based ellipse recognition method, belonging to the technical field of pattern recognition. Said method comprises: in a deep learning training stage, first acquiring an image containing an ellipse, and constructing a training data set after labeling edges, then establishing a deep neural network, and using the training data set to perform iterative training on the deep neural network, so as to obtain a trained network; and in an edge tracking and ellipse recognition stage, inputting, into the trained network, an image to be recognized, outputting a predicted edge image by the network, and then recognizing the ellipse in the edge image by using an edge tracking algorithm, so as to complete the recognition.

Description

An ellipse recognition method based on deep learning

Cross references to related applications

This application is based on a Chinese patent application filed with the application number: 201910732288.4, and the filing date is August 9, 2019, and claims the priority of the Chinese patent application. The entire content of the Chinese patent application is hereby incorporated by reference into this application.

Technical field

This application relates to the technical field of pattern recognition, and specifically to an ellipse recognition method based on deep learning.

Background technique

Since the ellipse is one of the most common geometric figures, ellipse recognition has become the basis of many computer vision applications. For example, ellipse recognition can be used in surface inspection, camera calibration, object distinction, eye capture, road sign recognition and classification. Therefore, it is very important to recognize the ellipse robustly and stably from the image.

The existing ellipse recognition algorithms are roughly divided into two categories, one is the ellipse recognition algorithm based on Hough transform, and the other is the ellipse recognition algorithm based on edge tracking. The ellipse recognition algorithm based on Hough transform is not only inefficient, but also inaccurate, which makes it difficult to be used in an industrial generation environment. The existing ellipse recognition algorithms based on edge tracking are all based on the image gradient to identify the edge, and then based on the edge to identify the ellipse. For pictures in the industrial environment, because the image gradient noise is more, the image edge cannot be well recognized only based on the image gradient. In addition, none of the existing ellipse recognition algorithms based on edge tracking can well recognize small ellipses (ellipses whose semi-major axis is less than 10 pixels in the image).

Summary of the invention

The purpose of this application is to overcome the shortcomings of the prior art and propose an ellipse recognition method based on deep learning. This application uses deep learning algorithms to identify edge images, which can overcome the problem of image gradient noise. Edge tracking algorithms are used to identify ellipses on edge images, which optimizes the recognition of small ellipses, and can achieve high precision and high recall. High application value.

This application proposes an ellipse recognition method based on deep learning, which is characterized in that the method is divided into two stages of deep learning training and edge tracking to identify ellipses, including the following steps:

1) Deep learning training stage; the specific steps are as follows:

1-1) Collect M images containing ellipses as training input images, manually label all the edge information contained in each training input image, and obtain the corresponding edge image as the training target image after labeling; each training input image and Form a training sample corresponding to the training target image, and all training samples form a training data set;

1-2) Build a deep neural network; the network has 14 layers of neurons, the first 7 layers of the network are compressed layers, the input image size of each layer of the compressed layer is reduced in turn in a half way, and the next 7 layers are generating layers, generating layers The input image size of each layer is expanded in a doubling manner. The output image size of the 7th generation layer is the same as the input image size; the network outputs a total of 8 images, assuming the input image size is W×W, then 8 output The dimensions of the images O _d0 , O _d1 , O _d2 , O _d3 , O _u3 , O _u2 , O _u1 , O _u0 are respectively W×W, 0.5W×0.5W, 0.25W×0.25W, 0.125W×0.125W, 0.125 W×0.125W, 0.25W×0.25W, 0.5W×0.5W, W×W; among them, O _dj,i is the output of the compression layer, O _uj,i is the output of the generation layer, j=0,1,2 ,3 represents the number of times the image is halved compared to the original image size;

1-3) Use the training data set to perform iterative training on the deep neural network to obtain the trained deep neural network; the specific steps are as follows:

1-3-1) Randomly initialize the parameters of the deep neural network established in 1-2), and use the initial deep neural network as the current neural network;

1-3-2) Let i=1;

1-3-3) Use the current deep neural network to predict the output edge image corresponding to each training input image in the training data set; record the i-th training input image as X _0,i , the 8 output edge images predicted by the deep neural network Respectively O _dj,i and O _uj,i , j=0,1,2,3;

1-3-4) the first web training input image i X _{0, i} corresponding to the target image training Y _{0, i} consecutive binary down-sampled _{Y j, i, j = 0,1,2,3} , where Y _{0, The} size of _i is W×W, then the size of Y _j,i is 0.5 ^j W×0.5 ^j W;

1-3-5) Using the output edge image O _dj,i , O _uj,i and the training target image Y _{j,i to} construct a multi-layer error _Li , the expression is as follows:

1-3-6) Use the gradient descent algorithm to update the parameters of the current deep neural network by minimizing L _i to obtain the updated current deep neural network;

1-3-7) Let i=i+1, and then go back to step 1-3-3), use gradient descent and other algorithms to iteratively train the current deep neural network by minimizing L _i until all training samples are traversed After that, return to step 1-3-2) for the next iteration of training until the average error

There is no decline in 5 consecutive rounds of iterative training, where N is the number of training samples; the deep neural network when _{Lavg is the} smallest is the deep neural network that has been trained;

2) Edge tracking to identify the ellipse; the specific steps are as follows:

2-1) Acquire an image X _a0 (input the image into step 1) of the trained deep neural network arbitrarily, and obtain the output image O _{a0 of} the 7th generation layer of the network;

2-2) Perform morphological slimming processing on the image O _a0 obtained in step 2-1) to obtain an edge image with a width of 1 pixel;

2-3) Perform edge tracking on the edge image with a width of 1 pixel in step 2-2); the specific steps are as follows:

2-3-1) Randomly select 2-2) Obtain an unused pixel in the edge image as a seed pixel, the seed pixel contains no more than two 8-connected neighbors;

2-3-2) Initialize an empty curve without any pixels as the current edge curve, select any 8-connected neighbor of the seed pixel as the boundary pixel, and add the boundary pixel to the current edge curve, all add the current edge curve The pixels are used pixels;

2-3-3) Judge the boundary pixels:

If the boundary pixel contains more than two or less than two 8-connected neighbors, the edge tracking of the connection direction between the seed pixel and the boundary pixel is ended, and step 2-3-5); if the boundary pixel contains two 8-connected neighbors , The 8-connected neighbors of the two 8-connected neighbors of the boundary pixel that are not on the current edge curve are regarded as the new boundary pixels, and the current edge curve is added to step 2-3-4);

2-3-4) Repeat steps 2-3-3) until the edge tracking in the direction of the connection between the seed pixel and the boundary pixel ends, and then proceed to step 2-3-5);

2-3-5) If the seed pixel has only one 8-connected neighbor, then use the current edge curve that has completed edge tracking in step 2-3-4) as the edge curve of the seed pixel, and proceed to step 2-4); if the seed pixel If there are two 8-connected neighbors, select another 8-connected neighbor of the seed pixel as the new boundary pixel, repeat 2-3-2) to 2-3-4), initialize a new current edge curve, and complete the seed pixel Edge tracking in the direction of the connection with the new boundary pixel, and finally the two current edge curves that have completed edge tracking in steps 2-3-4) and 2-3-5) are spliced into an edge curve of the seed pixel;

2-4) Randomly select a new unused pixel from the edge image in step 2-2), and repeat steps 2-3) until all unused pixels get the corresponding edge curve;

2-5) Scan each edge curve obtained in step 2-4), and disconnect the edge curve where the curvature of each edge curve exceeds the set curvature threshold;

2-6) Delete the edge curves whose straightness exceeds the set straightness threshold from the edge curves processed in step 2-5);

2-7) For each end point of each edge curve processed in step 2-6), calculate the edge curve connected to the end point to obtain a smooth curve, and sort the calculated edge curves from large to small in length, Take the edge curve with the largest length as the neighbor curve of the end point;

2-8) Randomly select an edge curve from the edge curves processed in step 2-6), splice the neighboring curve of the end along any end of the edge curve, and then continue with the end of the neighboring curve that is not used for splicing Splice the corresponding neighbor curves, and then go down in turn until the splicing is impossible, save the selected edge curve and the curve group obtained after each splicing respectively, and obtain the intermediate process curve group corresponding to the edge curve and the final splicing result curve group;

2-9) Repeat steps 2-8) until the intermediate process curve group and the final splicing result curve group corresponding to all the edge curves are obtained;

2-10) Calculate the rotation angle of each curve group obtained in step 2-9). The calculation method is to start from one end of the curve group and accumulate the rotation angle along the curve until the other end of the curve group. The rotation angle is recorded as the rotation angle of the curve group;

Use the rotation angle of each curve group to filter all curve groups: if the rotation angle of the curve group is less than 150 degrees or greater than 400 degrees, the curve group is an illegal combination, and the curve group is deleted;

2-11) Fit the corresponding ellipse to each curve group after screening in step 2-10), calculate the distance from the pixel on the curve group to the fitted corresponding ellipse and judge: if the distance to the fitted corresponding ellipse If the number of pixels with a distance of less than one pixel to all pixels is lower than the set ratio threshold, the ellipse will be deleted, otherwise the ellipse will be retained;

2-12) For all ellipses processed in 2-11), calculate the ratio of the number of pixels whose distance from the pixel to the ellipse in the corresponding curve group is less than one pixel distance divided by the circumference of the ellipse, and sort all ellipses in descending order of the ratio. ；

2-13) Select the ellipses sorted in step 2-12) in order and make a judgment: if the edge curves used by the selected ellipse are not consumed, the ellipse is retained and the edges used to fit the ellipse are consumed at the same time Curve; otherwise delete the ellipse; after all ellipses are judged, the ellipse finally retained is the result of step 2-1) to obtain the ellipse recognition corresponding to the image.

The features and beneficial effects of this application are:

In this application, a deep neural network can be trained to recognize edge images in a high-noise environment, and then edge tracking algorithms can be used to identify ellipses from edge images. This application is suitable for identifying ellipses in an industrial environment, can well deal with the problems of high noise in industrial environment images and many small ellipses, and has high recognition result accuracy and recall rate, and has high application value.

Description of the drawings

Figure 1 is an overall flowchart of an embodiment of the present application;

Fig. 2 is a schematic diagram of a deep neural network structure according to an embodiment of the present application.

detailed description

This application proposes an ellipse recognition method based on deep learning. The application will be further described in detail below with reference to the drawings and specific embodiments.

This application proposes an ellipse recognition method based on deep learning, which is divided into two stages of deep learning training and edge tracking to identify ellipses. The overall process is shown in Figure 1, including the following steps:

1) Deep learning training stage; the specific steps are as follows:

1-1) Collect images containing ellipses as training input images. The more the number, the better. In this embodiment, 81 images containing ellipses collected by the network are used as training input images. The size ranges from 200 pixels to 1000 pixels. Manually label all the edge information (not limited to ellipse) contained in each training input image, and get the corresponding edge image as the training target image after labeling; each training input image and the corresponding training target image form a training sample, all training The samples constitute the training data set;

1-2) Construct a deep neural network; the input of the deep neural network is an image containing an ellipse, and the output is an edge image corresponding to the input image. The network has 14 layers of neurons. The first 7 layers of the network are compressed layers. The input image size of each layer of the compressed layer is reduced in a half way. The next 7 layers are generating layers. The input image size of each layer of the generating layer is doubled. The method is successively expanded, and the output image size of the seventh generation layer is the same as the input image size. Each layer is connected to an output module except the merged layer (the layer represented by the two rectangles superimposed in Figure 2). The network outputs a total of 8 images. Assuming the input image size is W×W, then 8 output images O _d0 , O _d1 , O _d2 , O _d3 , O _u3 , O _u2 , O _u1 , O _u0 are respectively W×W, 0.5W×0.5W, 0.25W×0.25W, 0.125W×0.125W, 0.125W×0.125 W, 0.25W×0.25W, 0.5W×0.5W, W×W;

FIG. 2 is a schematic diagram of a result of a deep neural network according to an embodiment of the application. Among them, 1 is the input image, and 2 is the output image of each layer. Among them, Pool,/2 is the half-pooling layer, Conv is the convolutional layer, Conv,/2 is the half-convolutional layer, Trans,X2 is the double generation layer, and the two graphics connected by Same represent the same intermediate variable; X _{0 ,i} is the i-th input image, X _1,i , X _2,i , X _3,i are the images reduced to one-half, one-quarter, and one-eighth of X _0,i respectively; Z _{dj ,i} is the intermediate variable of the compression layer, Z _uj,i is the intermediate variable of the generation layer, O _dj,i is the output of the compression layer, O _uj,i is the output of the generation layer, where j=0,1,2,3 , Represents the number of times the image is halved compared to the original input image size;

1-3-1) Randomly initialize the parameters of the deep neural network established in 1-2), or use some mature initialization algorithms. Use the initial deep neural network as the current neural network;

1-3-2) Let i=1;

1-3-3) Using the current deep neural network to predict the output edge image corresponding to each training input image in the training data set, where each training input image corresponds to 8 output edge images of different sizes. Denote the i-th training input image as X _0,i , and the 8 output edge images predicted by the deep neural network are O _dj,i (j=0,1,2,3) and O _uj,i (j=0 ,1,2,3), the image size with subscript j is 0.5 ^j of the original image;

1-3-4) the first web training input image i X _{0, i} corresponding to the target image training Y _{0, i} consecutive binary down-sampled _{Y j, i (j = 1,2,3} ), where j represents a binary drop The number of samplings, where Y _0,i size is W×W, then Y _j,i size is 0.5 ^j W×0.5 ^j W;

1-3-6) Use the gradient descent algorithm to update the parameters of the current deep neural network by minimizing L _i to obtain the updated current deep neural network. Batch update can be set according to the memory situation. This embodiment updates each image once;

1-3-7) Let i=i+1, and then go back to step 1-3-3), use gradient descent and other algorithms to iteratively train the current deep neural network by minimizing L _i until all training samples are traversed After that, return to step 1-3-2) for the next iteration of training until the average error is reached

(Where N is the number of training samples in the training data set) 5 consecutive rounds of iterative training no longer drop, take the deep neural network when _{Lavg is the} smallest as the trained deep neural network;

2) Edge tracking to identify the ellipse; the specific steps are as follows:

2-1) Obtain an image X _a0 arbitrarily, and input the image into the trained deep neural network of step 1) to obtain the output image O _{a0 of} the 7th generation layer of the network; where the image X _a0 can be obtained in any way Image;

2-2) Morphologically thinning the image O _a0 obtained in step 2-1) to obtain an edge image with a width of 1 pixel;

2-3-1) Randomly select 2-2) Obtain an unused pixel in the edge image (all pixels of the edge image are not used in the initial state) as the seed pixel, and the seed pixel must satisfy No more than two 8-connected neighbors, 8-connected neighbors are pixels that share an edge or a corner with the pixel;

2-3-2) Initialize an empty curve without any pixels as the current edge curve, select any 8-connected neighbor of the seed pixel as the boundary pixel, and add the boundary pixel to the current edge curve (the curve is composed of pixels, Adding is to put pixels in the curve, which is equivalent to the curve lengthening by one pixel), all pixels added to the current edge curve are considered used;

2-3-3) Judge the boundary pixels:

If the boundary pixel contains more than two or less than two 8-connected neighbors, then end the edge tracking in the direction of the connection between the seed pixel and the boundary pixel, and proceed to step 2-3-5); if the boundary pixel contains exactly two 8-connected neighbors Neighbors, the 8-connected neighbors of the two 8-connected neighbors of the boundary pixel that are not on the current edge curve are used as the new boundary pixels, and the current edge curve is added to step 2-3-4);

2-4) Randomly select a new unused pixel from the edge image in step 2-2), repeat steps 2-3), until all unused pixels have corresponding edge curves, so far, a The edge curve of the series;

2-5) Scan each edge curve obtained in step 2-4), and disconnect the edge curve where the curvature exceeds the set curvature threshold in each edge curve to obtain multiple edge curves. The specific curvature threshold is selected according to the application scenario (In this embodiment, the threshold is twice the average value);

2-6) Delete edge curves whose straightness exceeds the set straightness threshold from the edge curves processed in step 2-5), and the straightness threshold is adjusted according to specific application scenarios (0.9 in this embodiment);

2-7) For each end point of each edge curve processed in step 2-6), calculate the edge curve that can be connected to the end point to obtain a smooth curve, and sort the calculated edge curves from large to small in length , Take the edge curve with the largest length as the neighbor curve of the end point;

2-8) Randomly select an edge curve from the edge curves processed in step 2-6), splice the neighboring curve of the end along any end of the edge curve, and then continue with the end of the neighboring curve that is not used for splicing Splice the corresponding neighbor curves, and then go down in turn until the splicing is impossible, save the selected edge curve and the curve group obtained after each splicing respectively, and obtain the intermediate process curve group corresponding to the edge curve and the final splicing result curve group; For example, if the final splicing result curve group [P ₁ ,P ₂ ,P ₃ ] of a certain edge curve exists, then the intermediate process [P ₁ ] and [P ₁ ,P ₂ ] are also retained, where P ₁ is randomly selected An edge curve, P ₂ is the neighbor curve of one end of P ₁ , and P ₃ is the neighbor curve corresponding to the end of P _{2 that} is not used for splicing.

2-9) Repeat steps 2-8) until all the edge curves correspond to the intermediate process curve group and the final splicing result curve group; so far, a series of curve combinations are obtained;

2-10) Calculate the rotation angle of each curve group obtained in step 2-9). The calculation method is to start from one end of the curve group and accumulate the rotation angle along the curve until the other end of the combination, the cumulative rotation The angle is recorded as the rotation angle of the curve group;

Use the rotation angle of each curve group to filter all curve groups: if the rotation angle of the curve group is less than 150 degrees or greater than 400 degrees, the curve group is considered to be an illegal combination, and the curve group is deleted;

2-11) Fit the corresponding ellipse to each curve group after screening in step 2-10), and calculate the distance from the pixel on the curve group to the fitted corresponding ellipse, where the distance to the fitted corresponding ellipse is less than one pixel The ratio of the number of distance pixels to the number of all pixels shall not be lower than the set ratio threshold. If the number of pixels less than one pixel distance to the number of all pixels is lower than the set ratio threshold, the ellipse will be deleted, otherwise the ellipse will be retained (specific threshold The selection is determined according to actual application scenarios, and the ratio threshold value in this embodiment is 0.5);

2-12) For all ellipses processed in 2-11), calculate the ratio of the number of pixels whose distance from the pixel to the ellipse in the corresponding curve group is less than one pixel distance divided by the circumference of the ellipse, and sort all by the ratio from largest to smallest oval;

2-13) Select the sorted ellipses in step 2-12) in order and make a judgment: if none of the edge curves used by the selected ellipse are consumed, the ellipse is retained and the ellipse used for fitting the ellipse is consumed at the same time Edge curve; otherwise delete the ellipse (that is, if all or part of the edge curve fitting the ellipse has been consumed, delete the ellipse); after all ellipses are judged, the final ellipse remaining is step 2-1) Get the image The result of corresponding ellipse recognition.

This application uses a deep neural network to recognize edge images, and uses an edge tracking algorithm to recognize ellipses, and realizes ellipse recognition based on deep learning.

Claims

An ellipse recognition method based on deep learning, characterized in that the method is divided into two stages of deep learning training and edge tracking to identify ellipse, including the following steps:

1) Deep learning training stage; the specific steps are as follows:

1-1) Collect M images containing ellipses as training input images, manually label all the edge information contained in each training input image, and obtain the corresponding edge image as the training target image after labeling; each training input image and Form a training sample corresponding to the training target image, and all training samples form a training data set;

1-2) Build a deep neural network; the network has 14 layers of neurons, the first 7 layers of the network are compressed layers, the input image size of each layer of the compressed layer is reduced in turn in a half way, and the next 7 layers are generating layers, generating layers The input image size of each layer is expanded in a doubling manner. The output image size of the 7th generation layer is the same as the input image size; the network outputs a total of 8 images, assuming the input image size is W×W, then 8 output The dimensions of the images O d0 , O d1 , O d2 , O d3 , O u3 , O u2 , O u1 , O u0 are respectively W×W, 0.5W×0.5W, 0.25W×0.25W, 0.125W×0.125W, 0.125 W×0.125W, 0.25W×0.25W, 0.5W×0.5W, W×W; among them, O dj,i is the output of the compression layer, O uj,i is the output of the generation layer, j=0,1,2 ,3 represents the number of times the image is halved compared to the original image size;

1-3) Use the training data set to perform iterative training on the deep neural network to obtain the trained deep neural network; the specific steps are as follows:

1-3-1) Randomly initialize the parameters of the deep neural network established in 1-2), and use the initial deep neural network as the current neural network;

1-3-2) Let i=1;

1-3-3) Use the current deep neural network to predict the output edge image corresponding to each training input image in the training data set; record the i-th training input image as X 0,i , the 8 output edge images predicted by the deep neural network Respectively O dj,i and O uj,i , j=0,1,2,3;

1-3-4) the first web training input image i X 0, i corresponding to the target image training Y 0, i consecutive binary down-sampled Y j, i, j = 0,1,2,3 , where Y 0, The size of i is W×W, then the size of Y j,i is 0.5 j W×0.5 j W;

1-3-5) Using the output edge image O dj,i , O uj,i and the training target image Y j,i to construct a multi-layer error Li , the expression is as follows:

1-3-6) Use the gradient descent algorithm to update the parameters of the current deep neural network by minimizing L i to obtain the updated current deep neural network;

1-3-7) Let i=i+1, and then go back to step 1-3-3), use gradient descent and other algorithms to iteratively train the current deep neural network by minimizing L i until all training samples are traversed After that, return to step 1-3-2) for the next iteration of training until the average error
There is no decline in 5 consecutive rounds of iterative training, where N is the number of training samples; the deep neural network when Lavg is the smallest is the deep neural network that has been trained;

2) Edge tracking to identify the ellipse; the specific steps are as follows:

2-1) Acquire an image X a0 (input the image into step 1) of the trained deep neural network arbitrarily, and obtain the output image O a0 of the 7th generation layer of the network;

2-2) Perform morphological slimming processing on the image O a0 obtained in step 2-1) to obtain an edge image with a width of 1 pixel;

2-3) Perform edge tracking on the edge image with a width of 1 pixel in step 2-2); the specific steps are as follows:

2-3-1) Randomly select 2-2) Obtain an unused pixel in the edge image as a seed pixel, which contains no more than two 8-connected neighbors;

2-3-2) Initialize an empty curve without any pixels as the current edge curve, select any 8-connected neighbor of the seed pixel as the boundary pixel, and add the boundary pixel to the current edge curve, all add the current edge curve The pixels are used pixels;

2-3-3) Judge the boundary pixels:

If the boundary pixel contains more than two or less than two 8-connected neighbors, the edge tracking of the connection direction between the seed pixel and the boundary pixel is ended, and step 2-3-5); if the boundary pixel contains two 8-connected neighbors , The 8-connected neighbor of the two 8-connected neighbors of the boundary pixel that is not on the current edge curve is taken as the new boundary pixel, and the current edge curve is added to step 2-3-4);

2-3-4) Repeat step 2-3-3) until the edge tracking in the direction of the connection between the seed pixel and the boundary pixel is finished, and then enter step 2-3-5);

2-3-5) If the seed pixel has only one 8-connected neighbor, then use the current edge curve that has completed edge tracking in step 2-3-4) as the edge curve of the seed pixel, and proceed to step 2-4); if the seed pixel If there are two 8-connected neighbors, select another 8-connected neighbor of the seed pixel as the new boundary pixel, repeat 2-3-2) to 2-3-4), initialize a new current edge curve, and complete the seed pixel Edge tracking in the direction of the connection with the new boundary pixel, and finally the two current edge curves that have completed edge tracking in steps 2-3-4) and 2-3-5) are spliced into an edge curve of the seed pixel;

2-4) Randomly select a new unused pixel from the edge image in step 2-2), and repeat steps 2-3) until all unused pixels get the corresponding edge curve;

2-5) Scan each edge curve obtained in step 2-4), and disconnect the edge curve where the curvature of each edge curve exceeds the set curvature threshold;

2-6) Delete the edge curves whose straightness exceeds the set straightness threshold from the edge curves processed in step 2-5);

2-7) For each end point of each edge curve processed in step 2-6), calculate the edge curve connected to the end point to obtain a smooth curve, and sort the calculated edge curves from large to small in length, Take the edge curve with the largest length as the neighbor curve of the end point;

2-8) Randomly select an edge curve from the edge curves processed in step 2-6), splice the neighboring curve of the end along any end of the edge curve, and then continue with the end of the neighboring curve that is not used for splicing Splice the corresponding neighbor curves, and then go down in turn until the splicing is impossible, save the selected edge curve and the curve group obtained after each splicing respectively, and obtain the intermediate process curve group corresponding to the edge curve and the final splicing result curve group;

2-9) Repeat steps 2-8) until the intermediate process curve group and the final splicing result curve group corresponding to all the edge curves are obtained;

2-10) Calculate the rotation angle of each curve group obtained in step 2-9). The calculation method is to start from one end of the curve group and accumulate the rotation angle along the curve until the other end of the curve group. The rotation angle is recorded as the rotation angle of the curve group;

Use the rotation angle of each curve group to filter all curve groups: if the rotation angle of the curve group is less than 150 degrees or greater than 400 degrees, the curve group is an illegal combination, and the curve group is deleted;

2-11) Fit the corresponding ellipse to each curve group after screening in step 2-10), calculate the distance from the pixel on the curve group to the fitted corresponding ellipse and judge: if the distance to the fitted corresponding ellipse If the number of pixels with a distance of less than one pixel to all pixels is lower than the set ratio threshold, the ellipse will be deleted, otherwise the ellipse will be retained;

2-12) For all ellipses processed in 2-11), calculate the ratio of the number of pixels whose distance from the pixel to the ellipse in the corresponding curve group is less than one pixel distance divided by the circumference of the ellipse, and sort all ellipses in descending order of the ratio. ；

2-13) Select the ellipses sorted in step 2-12) in order and make a judgment: if the edge curves used by the selected ellipse are not consumed, the ellipse is retained and the edges used to fit the ellipse are consumed at the same time Curve; otherwise delete the ellipse; after all ellipses are judged, the ellipse finally retained is the result of step 2-1) to obtain the ellipse recognition corresponding to the image.