Edge feature extraction method based on deep learning target recognition

Abstract

The application discloses an edge feature extraction method based on deep learning target identification, acquires a picture to be detected, determines a detection range of the picture to be detected, determines a gray scale change inflection point of the detection range, determines a training range according to the detection range, outputs boundary feature points through a CNN model, determines the gray scale change inflection points in a threshold radius range of the boundary feature points as feature points, and fits the boundary feature points and the feature points to form an edge line. The application combines the innovative deep learning detection and the screening function of the boundary feature points in the boundary algorithm, preliminarily positions through the CNN model, improves the stability in a noise environment, determines the approximate range of the boundary, and positions the accurate position in the approximate range through the boundary algorithm. The application still has the ability to maintain stable detection performance when abnormal data, noise or environmental changes occur, has the precision advantage of boundary extraction, can optimize the boundary position, the precision can reach the sub-pixel level, and the robustness and stability of detection are greatly improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of artificial intelligence, and in particular relates to an edge feature extraction method based on deep learning target recognition. BACKGROUND

[0002] In modern industrial manufacturing processes, the size and shape precision requirements of products are increasingly high. Due to its high precision and high efficiency in measurement, the two-dimensional image measuring instrument is widely used in mechanical processing, electronic components, mold manufacturing and other fields. The two-dimensional image measuring instrument obtains the image of the measured object through a high-resolution camera, and processes and analyzes it using special software, thereby realizing fast and accurate size measurement and shape analysis. However, when measuring some workpieces with high surface brightness, the two-dimensional image measuring instrument faces a technical challenge: due to the unclear boundary of the surface light, it is difficult to automatically identify the boundary. For example, when measuring some high-reflective metal surfaces or products with complex curved surfaces, the reflection and scattering of light will make the boundary blurred, and the automatic edge extraction algorithm of the image measuring instrument cannot accurately identify the true boundary. In this case, the automatic measurement program cannot work normally, and accurate measurement data cannot be obtained, thereby affecting product quality control and production efficiency. In order to solve this problem, the current solution is to use the method of manually picking up the boundary. The operator needs to observe the image captured by the image measuring instrument and manually mark the boundary position on the software interface. Although this method can ensure the accuracy of boundary identification, it greatly reduces the measurement efficiency, increases the labor cost, and has high skill requirements for the operator. In addition, manual boundary picking is also easily affected by subjective factors, and different operators may get different measurement results, resulting in poor repeatability and consistency of measurement data. Therefore, how to improve the automatic boundary recognition capability of the two-dimensional image measuring instrument when measuring workpieces with high surface brightness, and reduce manual intervention, has become an important direction of current technical development.

[0003] The above information disclosed in this BACKGROUND section is only for the purpose of increasing the understanding of the background of the application, and thus, it can include information that is not prior art to the present application. SUMMARY

[0004] The present application proposes an edge feature extraction method based on deep learning target recognition to solve the technical problems of poor automatic boundary recognition capability of the two-dimensional image measuring instrument, low measurement efficiency, high cost and low measurement accuracy caused by manual intervention.

[0005] To achieve the above-mentioned application / design purposes, the present application adopts the following technical solutions:

[0006] An edge feature extraction method based on deep learning target recognition, the method comprising:

[0007] obtaining a picture to be detected, determining a detection range of the picture to be detected;

[0008] determining a gray level change inflection point of the detection range, determining a training range according to the detection range, and outputting boundary feature points by a CNN model for the training range;

[0009] determining the gray level change inflection point within a threshold radius range of the boundary feature points as a feature point;

[0010] fitting the boundary feature points and the feature points to form an edge line.

[0011] The method for determining a training range according to the detection range in the edge feature extraction method based on deep learning target recognition as described above is:

[0012] annotating edges in the detection range, and determining the number N of annotation rectangular boxes to be generated in the detection range;

[0013] calculating the coordinates of the center points of the annotation rectangular boxes as the coordinates of equidistant points, and the coordinates of the i-th equidistant point are ;

[0014] determining the coordinates C (Cx, Cy) and D (Dx, Dy) of the opposite angles of the training range through the coordinates of the detection range;

[0015] Cx = min(x), Cy = min(y)); Dx = max(x), Dy = max(y));

[0016] annotating and converting according to the coordinates of the equidistant points and the training range, and inputting the converted annotation into the CNN model.

[0017] The edge feature extraction method based on deep learning target recognition as described above, the annotation includes line annotation:

[0018] determining the coordinates (Ax, Ay) and (Bx, By) of two end points A and B of a straight line;

[0019] calculating the length of the straight line ;

[0020] calculating the number N of annotation rectangular boxes to be generated,

[0021] Boxsize is the length and width of the annotation rectangular box determined in advance, and the overlap rate between the boxes is set , The range of is [0, 1);

[0022] The coordinates of the center point of the labeled rectangular frame are calculated as the coordinates of the equidistant points, and the coordinates of the i-th equidistant point are Then:

[0023] ;

[0024] Where i is from 0 to N-1.

[0025] The edge feature extraction method based on deep learning target recognition as described above, the label includes arc labeling:

[0026] Determine the center coordinates (h, k) of the arc circle, the central angle θ of the arc, and the radius r of the arc circle;

[0027] Calculate the length of the arc ;

[0028] Calculate the number N of labeled rectangular frames needed to be generated, ;

[0029] Where Boxsize is the length and width of the labeled rectangular frame determined in advance, and the overlap rate between the frames is set , The range of is [0, 1); calculate the bisector angle Angle=θ / N;

[0030] Calculate the coordinates of the center point of the labeled rectangular frame as the coordinates of the equidistant points, and the coordinates of the i-th equidistant point are Then:

[0031] ;

[0032] Where r is the radius, is the starting angle, and i is 0 to N-1.

[0033] The edge feature extraction method based on deep learning target recognition as described above, the label includes circle labeling:

[0034] Determine the center coordinates (h, k) of the circle and the radius r of the circle;

[0035] Calculate the circumference L=2πr of the circle;

[0036] Calculate the number N of labeled rectangular frames needed to be generated, ;

[0037] Where Boxsize is the length and width of the labeled rectangular frame determined in advance, and the overlap rate between the frames is set , The range of is [0, 1);

[0038] Calculate the bisector angle Angle=360° / N;

[0039] The coordinates of the center point of the labeled rectangular frame are calculated as the coordinates of the equidistant points, and the coordinates of the i-th equidistant point are Then:

[0040] ;

[0041] where i is 0 to N-1.

[0042] In the edge feature extraction method based on deep learning target recognition as described above, the residual block of the CNN model is composed of two convolution layers, the first convolution layer uses a 1x1 convolution kernel, and the second convolution layer uses a 3x3 convolution kernel, the output of the residual block is the direct addition of the outputs of the two convolution layers and the input, and the 3x3 convolution layer with a step of 2 is down-sampled;

[0043] Loss calculation during backpropagation in model training ;

[0044] where, and is the loss of the center point of the prediction frame;

[0045] is the confidence loss, including the confidence loss of foreground and background;

[0046] Fixed anchor frame size: set all anchor frames to the same value in the configuration file;

[0047] The relative coordinates of the prediction frame used in the model training process are (tx, ty), and (tx, ty) is the displacement of the center coordinates of the prediction frame relative to the coordinates of the top-left corner of the grid (Cx, Cy).

[0048] In the edge feature extraction method based on deep learning target recognition as described above, the method for determining the gray level change inflection point of the detection range comprises:

[0049] The detection range of the image to be detected is subjected to gray scale processing to obtain a gray scale image, the gray level change rate of the gray scale image in the horizontal direction and the vertical direction is calculated, and the gray level change inflection point is determined by analyzing the local extreme points of the gray level change rate.

[0050] In the edge feature extraction method based on deep learning target recognition as described above, for the gray level change inflection point, the T value is used to filter the noise points, and the direction is the maximum difference direction.

[0051] In the edge feature extraction method based on deep learning target recognition as described above, the boundary feature points and feature points are fitted to form an edge line using the least squares method.

[0052] Compared with the prior art, the advantages and positive effects of the present application are: the edge feature extraction method based on deep learning target recognition of the present application comprises: acquiring a picture to be detected, determining a detection range (blue frame) of the picture to be detected; determining a gray scale change inflection point of the detection range; determining a training range (red frame) according to the detection range (blue frame), outputting boundary feature points of the training range (red frame) through a CNN model; determining the gray scale change inflection points within a threshold radius range of the boundary feature points as feature points; and fitting the boundary feature points and the feature points to form an edge line. The present application combines the screening function of the boundary feature points in the innovative deep learning detection and boundary algorithm, preliminarily positions through the CNN model, improves the stability in the noise environment, determines the approximate range of the boundary, and positions the accurate position in the approximate range through the boundary algorithm. On the one hand, the CNN model has the ability to maintain stable detection performance when the abnormal data, noise or environment changes, and on the other hand, the boundary extraction has the precision advantage, can optimize the boundary position, the precision can reach the sub-pixel level, and the robustness and stability of the detection are greatly improved.

[0053] Other features and advantages of the present application will become more apparent after reading the specific embodiments of the present application in combination with the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0055] Figure 1 is a schematic diagram of line labeling of a specific embodiment of the present application.

[0056] Figure 2 is a schematic diagram of arc labeling of a specific embodiment of the present application.

[0057] Figure 3 is a schematic diagram of circle labeling of a specific embodiment of the present application.

[0058] Figure 4 is a CNN model diagram used in a specific embodiment of the present application.

[0059] Figure 5 is an effect diagram of filtering noise points of a specific embodiment of the present application.

[0060] Figure 6 is an original inflection point diagram of a specific embodiment of the present application.

[0061] Figure 7 is an inflection point diagram filtered by T value of a specific embodiment of the present application.

[0062] Figure 8 is a schematic diagram of the boundary feature point obtained by the CNN model of the embodiment of the present application.

[0063] Figure 9 is a boundary schematic diagram obtained by fitting of the embodiment of the present application.

[0064] Figure 10 is a flowchart of the boundary detection of the embodiment of the present application. DETAILED DESCRIPTION

[0065] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0066] In the description of the present application, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the purpose of facilitating the description of the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore, it cannot be understood as a limitation to the present application.

[0067] In the description of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms "mounting", "connecting", "connection" should be understood broadly, for example, it can be fixed connection, or detachable connection, or integrally connected. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances. In the description of the embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0068] The terms "first", "second" are only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features limited by "first", "second" can explicitly or implicitly include one or more of the features.

[0069] In the description of the present application, unless otherwise specified, the meaning of "multiple" is two or more.

[0070] A deep learning target recognition-based edge feature extraction method, comprising:

[0071] An image to be detected is acquired, and a detection range (blue frame) of the image to be detected is determined.

[0072] Determine the gray level change inflection point of the detection range; determine the training range (red frame) according to the detection range (blue frame), and output the boundary feature point through the CNN model for the training range;

[0073] Determine the gray level change inflection point within the threshold radius range of the boundary feature point as the feature point.

[0074] Fit the boundary feature point and the feature point to form an edge line.

[0075] As shown in Figure 10 The edge feature extraction method based on deep learning target recognition includes the following steps:

[0076] S1, acquire a picture to be detected.

[0077] S2, determine the detection range of the picture to be detected.

[0078] S3, determine the gray level change inflection point of the detection range; determine the training range according to the detection range, and output the boundary feature point through the CNN model for the training range.

[0079] There is no order in determining the gray level change inflection point and the boundary feature point in this step, and they can be performed simultaneously.

[0080] S4, determine the gray level change inflection point within the threshold radius range of the boundary feature point as the feature point.

[0081] S5, fit the boundary feature point and the feature point to form an edge line.

[0082] The method for determining the training range according to the detection range is as follows:

[0083] Label the edges in the detection range, and determine the number N of labeled rectangular frames that need to be generated in the detection range;

[0084] Calculate the coordinates of the center points of the labeled rectangular frames as the coordinates of the equidistant points, and the coordinates of the i-th equidistant point are ;

[0085] Determine the coordinates C (Cx, Cy) and D (Dx, Dy) of the opposite corners of the training range through the coordinates of the detection range;

[0086] Cx = min(x), Cy = min(y); Dx = max(x), Dy = max(y);

[0087] According to the coordinates of the equidistant points and the training range, perform label conversion, and input the converted label into the CNN model.

[0088] In the traditional CNN model, the labeling method is only rectangular box or polygon, which is not suitable for boundary labeling. Therefore, the line labeling, circle labeling and arc labeling are designed in the embodiment, and the three labeling methods are described as follows:

[0089] The line labeling is composed of two labels:

[0090] 1. Rotatable rectangular box (red box): used for setting the detection range.

[0091] 2. Line: used for marking the position of the boundary.

[0092] Creation method:

[0093] First, confirm the starting point of the straight line. Second, confirm the end point of the straight line. Third, confirm the detection range.

[0094] In the example of Figure 1 , the line labeling method includes:

[0095] Determine the coordinates of the two end points A and B of the straight line (Ax, Ay), (Bx, By);

[0096] Calculate the length of the straight line ;

[0097] Calculate the number N of labeling rectangular boxes (yellow boxes) to be generated, ;

[0098] Boxsize is the length and width of the labeling rectangular box determined in advance, which sets the overlap rate between the boxes , The range of is [0, 1].

[0099] Calculate the coordinates of the center point of the labeling rectangular box as the coordinates of the equidistant points, and the coordinates of the i-th equidistant point are , then:

[0100] ;

[0101] Where i is from 0 to N-1.

[0102] The arc labeling is composed of two labels:

[0103] 1. Partial circular ring area: used for setting the detection range of the arc.

[0104] 2. Arc: used for marking the boundary of the arc.

[0105] Creation method:

[0106] First, confirm the starting point of the arc. Second, confirm any point on the arc. Third, confirm the end point of the arc. Fourth, confirm the range of the arc training picture.

[0107] In Figure 2 Examples, the method for annotating an arc comprises:

[0108] determining the center coordinate (h, k) of the circle where the arc is located, the central angle θ of the arc, and the radius r of the circle where the arc is located;

[0109] calculating the length of the arc ;

[0110] calculating the number of N of the annotation rectangular boxes needed to be generated, ; wherein Boxsize is the length and width of the annotation rectangular box determined in advance, and the overlap rate between the boxes is set , The range of the overlap rate is [0, 1).

[0111] calculating the bisector angle Angle = θ / N;

[0112] calculating the coordinate of the center point of the annotation rectangular box as the coordinate of the equidistant point, and the coordinate of the i-th equidistant point is , then:

[0113] ;

[0114] wherein r is the radius, is the starting angle, the angle of the 2 yellow arrows in the figure, and i is 0 to N-1.

[0115] The circle annotation is composed of two annotations:

[0116] 1. Annulus region: used to set the detection range of the circle.

[0117] 2. Circle: used to mark the boundary of the circle.

[0118] Creation method:

[0119] The first three points, the three points confirming the boundary of the circle. The fourth point, confirming the detection range of the circle.

[0120] In Figure 3 Examples, the method for annotating an arc comprises:

[0121] determining the center coordinate (h, k) of the circle where the arc is located, the central angle θ of the arc, and the radius r of the circle where the arc is located;

[0122] calculating the circumference L = 2πr of the circle;

[0123] calculating the number of N of the annotation rectangular boxes needed to be generated, ;

[0124] wherein Boxsize is the length and width of the annotation rectangular box determined in advance, and the overlap rate between the boxes is set , The range of the overlap rate between the setting boxes is [0, 1), and the overlap rate is set;

[0125] Calculate the bisector angle Angle=0 / N;

[0126] Calculate the coordinates of the center point of the annotation rectangle as the coordinates of the equidistant points, and the coordinates of the i-th equidistant point are Then:

[0127] ;

[0128] Where i is 0 to N-1. Because the standard CNN model does not support edge recognition annotation, the above annotation needs to be converted into a general annotation rectangle supported by the CNN model. Before the CNN model outputs the boundary feature points, the annotation conversion of the annotation of the detection range needs to be performed, and the annotation conversion method includes:

[0129] Determine the number N of annotation rectangle boxes that need to be generated in the detection range;

[0130] Calculate the coordinates of the center point of the annotation rectangle as the coordinates of the equidistant points, and the coordinates of the i-th equidistant point are ;

[0131] Determine the coordinates C (Cx, Cy) and D (Dx, Dy) of the opposite angles of the training range through the coordinates of the detection range;

[0132] Cx=min(x), Cy=min(y)); Dx=max(x), Dy=max(y));

[0133] Annotation generation:

[0134] <object-class>: class id of the target object, numbered from 0;

[0135] x_center = (Pi(x) - Cx) / width w, <x_center> is the x coordinate of the center point of the labeled rectangular frame, normalized to the range of 0 to 1 between the image width;

[0136] y_center = (Pi(y) - Cy) / height h, <y_center> is the y coordinate of the center point of the labeled rectangular frame, normalized to the range of 0 to 1 between the image height;

[0137] <width>= Boxsize / width w, <width>To annotate the width of the rectangular box, the range normalized to the image width is between 0 and 1;

[0138] <height>= Box size / height h, <height>To annotate the height of the rectangular frame, the range normalized to the image height is between 0 and 1;

[0139] where width w = max x - min x; height h = max y - min y.

[0140] The network composition of the CNN model mainly consists of convolutional layers, residual blocks (ResidualBlock), and pooling layers.

[0141] The network is divided into the front section, the middle section, and the back section:

[0142] The front section is mainly composed of convolutional layers and max pooling layers, which are used to extract the initial features of the image.

[0143] The middle section is mainly composed of residual blocks, which enhance the feature extraction ability and accelerate the convergence through residual connections.

[0144] The back section outputs the final prediction results through the global average pooling layer and the fully connected layer in the classification task.

[0145] Residual block structure:

[0146] Each residual block of the CNN model is composed of two convolutional layers:

[0147] The first convolutional layer uses a 1x1 convolutional kernel to reduce the number of channels.

[0148] The second convolutional layer uses a 3x3 convolutional kernel for feature extraction.

[0149] The output of the residual block is the direct addition of the outputs of the two convolutional layers and the input, which can prevent gradient vanishing.

[0150] Downsampling strategy: Use a 3x3 convolutional layer with a stride of 2 for downsampling instead of traditional max pooling layers. This strategy not only increases the learning ability of the network but also improves the flexibility of feature extraction.

[0151] Output features: Output three different size feature maps (13x13, 26x26, 52x52) for multi-scale target detection.

[0152] During the model training process, the loss calculation method of backpropagation and the fixed anchor box size are optimized. In the configuration file, all anchor boxes are set to boxsize, which greatly speeds up the training speed.

[0153] Loss calculation of backpropagation ;

[0154] where, and are the losses of the predicted box center points;

[0155] is the confidence loss, including the foreground and background confidence loss;

[0156] The relative coordinates of the prediction box in the model training process are (tx, ty), and (tx, ty) is the displacement of the center coordinates of the prediction box relative to the coordinates (Cx, Cy) of the upper left corner of the grid.

[0157] The CNN model is trained and saved.

[0158] The method for determining the gray scale change inflection point of the detection range comprises:

[0159] The detection range of the to-be-detected picture is subjected to grayscale processing to obtain a grayscale image, and the gray scale change rate of the grayscale image in the horizontal direction and the vertical direction is calculated.

[0160] Specifically, the Sobel operator can be used to calculate the gray scale change rate of the image in the horizontal direction and the vertical direction. The calculation formula is as follows:

[0161]

[0162] The gray scale change inflection point is determined by analyzing the local extreme points of the gray scale change rate. The gray scale change inflection point refers to the points on the gray scale change rate curve, and the calculation thereof is obtained by first-order derivation, and the calculation formula is as follows:

[0163]

[0164] For the gray scale change inflection point, the T value and the direction are used to filter the noise points, the T value is the difference value of adjacent pixels, and the direction is the maximum difference direction.

[0165] The T value and the direction are used to filter the noise points, thereby increasing the turning points from decrease to increase or from increase to decrease, and the filtering effect is shown in Figure 5 .

[0166] The T value is the difference value of adjacent pixels, and the direction is the maximum difference direction. The threshold value refers to the difference value within a certain range, for example, the value between 10 and 50.

[0167] For example, the original inflection point diagram is shown in Figure 6 , and the inflection point diagram after filtering by the T value is shown in Figure 7 .

[0168] The above-mentioned CNN model is used to detect the boundary feature points in the training range, and the result shown in Figure 8 is obtained.

[0169] The gray scale change inflection points within the threshold radius range of the determined boundary feature points are set as the feature points. ​

[0170] Boundary feature points and feature points are fitted to form edge lines.

[0171] In this process, the boundary feature points and feature points are fitted using the least squares method to form the edge lines, and the final boundary effect is as follows: Figure 9 As shown.

[0172] The resulting curved edges better reflect the true contours of objects in the image, improving the effectiveness of edge recognition.

[0173] This implementation example has the following beneficial effects:

[0174] 1. Improve annotation efficiency.

[0175] Simplified annotation process: Traditional annotation methods usually require manual drawing of multiple rectangles of the same size for the boundary, while this embodiment only requires marking once to automatically convert multiple rectangles of the same size.

[0176] Reduce annotation time: The new annotation format is easy to create and can be automated, enabling the annotation of large amounts of data to be completed in a short time.

[0177] Improved annotation accuracy: The new annotation format reduces human error due to fewer operations, thereby improving the overall accuracy of annotations.

[0178] Therefore, the innovative deep learning annotation format not only significantly improves annotation efficiency but also enables automatic annotation.

[0179] 2. Improve detection accuracy and the ability to grasp complex boundaries:

[0180] Intelligent Inflection Points on Complex Boundaries: Traditional annotation methods typically filter boundary points based solely on information such as direction, intensity, and location. This can lead to errors when boundary features vary significantly, resulting in the inability to correctly identify the boundary. This embodiment utilizes a deep learning model to extract the distribution of feature points, replacing the fixed-parameter filtering of traditional methods and improving the stability of edge detection on complex boundaries.

[0181] The calculation accuracy of sub-pixel inflection points is higher: traditional deep learning outputs inflection points only at the pixel level, failing to reach the sub-pixel level. By using boundary feature points obtained through deep learning to filter inflection points acquired through traditional T-degree calculations, the boundary stability of edge detection at the sub-pixel level is improved.

[0182] Therefore, this embodiment significantly improves the robustness and stability of detection by combining innovative deep learning detection with the traditional edge-grabbing function of filtering mid-boundary feature points.

[0183] The above examples are only used to illustrate the technical solutions of the present application, but not to limit the present application; although the present application has been described in detail with reference to the foregoing examples, the technical solutions recorded in the foregoing examples can still be modified by those of ordinary skill in the art, or some technical features thereof can be replaced by equivalent replacements; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions claimed by the present application.< / height> < / height> < / width> < / width>

Claims

1. An edge feature extraction method based on deep learning target recognition, characterized in that, The method comprises: acquiring a picture to be detected, and determining a detection range of the picture to be detected; determining a gray level change inflection point of the detection range, determining a training range according to the detection range, and outputting a boundary feature point by a CNN model for the training range; determining a gray level change inflection point in a threshold radius range of the boundary feature point as a feature point; fitting the boundary feature point and the feature point to form an edge line, and the method for determining the training range according to the detection range is: labeling an edge in the detection range, and determining a number N of labeled rectangular frames to be generated in the detection range; The coordinates of the center point of the labeled rectangular frame are calculated as the coordinates of the isometric points, and the coordinates of the i-th isometric point are ; determining coordinates C (Cx, Cy) and D (Dx, Dy) of opposite angles of the training range through coordinates of the detection range; Cx = min(x), Cy = min(y); Dx = max(x), Dy = max(y); labeling conversion is performed according to the coordinates of the equidistant points and the training range, and the converted label is input into the CNN model. 2.The edge feature extraction method based on deep learning target recognition according to claim 1, characterized in that, The labeling includes line labeling: determining coordinates (Ax, Ay) and (Bx, By) of two end points A and B of a straight line; Computing the length of a straight line ; The calculation needs to generate N labeled rectangular frames, ; Boxsize is the length and width of the bounding box determined in advance, and the overlap rate between the boxes is set , The range of the overlap rate is [0, 1). The coordinates of the center point of the labeled rectangular frame are calculated as the coordinates of the isometric points, and the coordinates of the i-th isometric point are Then: ； ； wherein i is from 0 to N-1. 3.The edge feature extraction method based on deep learning target recognition according to claim 1, characterized in that, The labeling includes arc labeling: determining a center coordinate (h, k) of a circle in which an arc line is located, a central angle θ of the arc line, and a radius r of the circle in which the arc line is located; Computing the length of an arc ; The calculation needs to generate N labeled rectangular frames, ; Wherein, Boxsize is the length and width of the labeled rectangular frame determined in advance, and the overlap rate between the frames is set , The range of is [0, 1); the bisector angle Angle = θ / N is calculated. The coordinates of the center point of the labeled rectangular frame are calculated as the coordinates of the isometric points, and the coordinates of the i-th isometric point are Then: ； ； where r is the radius, is the start angle, i is 0 to N-1. 4.The edge feature extraction method based on deep learning target recognition according to claim 1, characterized in that, The labeling includes circle labeling: determining a center coordinate (h, k) of a circle and a radius r of the circle; calculating a circumference L = 2πr of the circle; The calculation needs to generate N labeled rectangular frames, ; wherein Boxsize is the length and width of the bounding box determined in advance, and the overlap rate between the boxes is set , is in the range [0, 1). calculating a bisector angle Angle = 360° / N; The coordinates of the center point of the labeled rectangular frame are calculated as the coordinates of the isometric points, and the coordinates of the i-th isometric point are Then: ； ； wherein i is from 0 to N-1. 5.The edge feature extraction method based on deep learning target recognition according to claim 1, characterized in that, The residual block of the CNN model is composed of two convolution layers, the first convolution layer uses a 1×1 convolution kernel, and the second convolution layer uses a 3×3 convolution kernel, the output of the residual block is the direct addition of the outputs of the two convolution layers and the input, and a 3×3 convolution layer with a step of 2 is used for down-sampling; Loss calculation for backpropagation during model training ; wherein, and is the loss of the prediction box center point; is a confidence loss, including foreground and background confidence losses; Fixing anchor box size: setting all anchor boxes to the same value in the configuration file; The relative coordinates of the prediction box used in the model training process are (tx, ty), and (tx, ty) is the displacement of the center coordinates of the prediction box relative to the coordinates (Cx, Cy) of the upper left corner of the grid. 6.The edge feature extraction method based on deep learning target recognition according to claim 1, characterized in that, The method for determining the gray level change inflection point of the detection range comprises: graying the detection range of the picture to be detected to obtain a gray image, calculating a gray level change rate of the gray image in a horizontal direction and a vertical direction, and determining a gray level change inflection point by analyzing local extreme points of the gray level change rate.

7. The edge feature extraction method based on deep learning target recognition according to claim 6, characterized in that, For the gray level change inflection point, T value and direction are used to filter noise points, T value is the difference value of adjacent pixels, and the direction is the maximum difference direction. 8.The edge feature extraction method based on deep learning target recognition according to claim 1, characterized in that, The boundary feature point and the feature point are fitted to form an edge line by using the least square method.

Images