Edge feature extraction method based on deep learning target recognition

Abstract

The invention discloses an edge feature extraction method based on deep learning target recognition, and the method comprises the steps: obtaining a to-be-detected picture, and determining the detection range of the to-be-detected picture; determining a gray scale change inflection point of the detection range; determining a training range according to the detection range, and outputting boundary feature points through a CNN model; determining gray scale change inflection points in a set threshold radius range of the boundary feature points as feature points; and fitting the boundary feature points and the feature points to form edge lines. Through combination of innovative deep learning detection and a boundary feature point screening function in a boundary algorithm, initial positioning is performed through a CNN model, stability in a noise environment is improved, an approximate range of a boundary is determined, and an accurate position in the approximate range is positioned through the boundary algorithm. According to the invention, the method can still maintain the stable detection performance when abnormal data, noise or environment changes, has the precision advantage of boundary extraction, can optimize the boundary position, enables the precision to reach a sub-pixel level, and greatly improves the detection robustness and stability.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically, it relates to an edge feature extraction method based on deep learning target recognition. Background Technology

[0002] In modern industrial manufacturing, the requirements for the dimensional and shape accuracy of products are becoming increasingly stringent. Two-dimensional (2D) image measuring instruments, due to their high-precision and high-efficiency measurement capabilities, are widely used in machining, electronic components, mold manufacturing, and other fields. 2D image measuring instruments acquire images of the object being measured using a high-resolution camera and process and analyze them using specialized software, thereby achieving rapid and accurate dimensional measurement and shape analysis. However, when measuring workpieces with high surface gloss, 2D image measuring instruments face a technical challenge: the unclear boundaries of surface light make automatic boundary identification difficult. For example, when measuring highly reflective metal surfaces or products with complex curved surfaces, light reflection and scattering blur the boundaries, making it difficult for the image measuring instrument's automatic edge extraction algorithm to accurately identify the true boundaries. In such cases, the automatic measurement program cannot function properly and cannot obtain accurate measurement data, thus affecting product quality control and production efficiency. To solve this problem, the current solution is to use a manual boundary picking method. Operators need to observe the image captured by the image measuring instrument and manually mark the boundary positions on the software interface. While this method ensures accurate boundary recognition, it significantly reduces measurement efficiency, increases labor costs, and requires highly skilled operators. Furthermore, manual boundary picking is susceptible to subjective factors; different operators may obtain different measurement results, leading to poor repeatability and consistency of the measurement data. Therefore, improving the automatic boundary recognition capability of 2D image measuring instruments when measuring workpieces with high surface gloss, and reducing manual intervention, has become an important direction for current technological development.

[0003] The information disclosed in this background section is only intended to enhance the understanding of the background technology of this application, and therefore may include prior art that is not known to those skilled in the art. Summary of the Invention

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

[0005] To achieve the above-mentioned invention / design objectives, the present invention adopts the following technical solution: An edge feature extraction method based on deep learning for target recognition, the method comprising: Obtain the image to be detected and determine the detection range of the image to be detected; Determine the grayscale inflection point of the detection range; determine the training range based on the detection range, and output boundary feature points for the training range using a CNN model; The grayscale inflection points within a set threshold radius of the boundary feature points are identified as feature points. The boundary feature points and feature points are fitted to form the edge line.

[0006] The edge feature extraction method for deep learning-based object recognition described above, wherein the method for determining the training range based on the detection range is as follows: The edges within the detection range are marked, and the number N of marked rectangles that need to be generated within the detection range is determined. The coordinates of the center point of the labeled rectangle are used as the coordinates of the equidistant points. The coordinates of the i-th equidistant point are... ; The coordinates of the diagonal of the training range, C(Cx, Cy) and D(Dx, Dy), are determined by the coordinates of the detection range. Cx=min(x), Cy=min(y)); Dx=max(x), Dy=max(y)); The labels are transformed based on the coordinates of the equidistant points and the training range, and the transformed labels are input into the CNN model.

[0007] As described above, in the edge feature extraction method for deep learning-based object recognition, the annotation includes line annotations: Determine the coordinates (Ax, Ay) and (Bx, By) of the two endpoints A and B of the line; Calculate the length of the line ; The calculation requires generating N labeled rectangles.

[0008] Boxsize is used to predetermine the length and width of the labeled rectangle, and sets the overlap ratio between the boxes. , The range is [0, 1); The coordinates of the center point of the labeled rectangle are used as the coordinates of the equidistant points. The coordinates of the i-th equidistant point are... ,but: ; Where i ranges from 0 to N-1.

[0009] As described above, in the edge feature extraction method for deep learning-based object recognition, the annotation includes arc annotation: Determine the coordinates (h,k) of the center of the circle containing the arc, the central angle θ of the arc, and the radius r of the circle containing the arc; Calculate the length of the arc ; The calculation requires generating N labeled rectangles. ; Boxsize is a pre-defined unit for the length and width of the labeled rectangle, and it sets the overlap ratio between the rectangles. , The range is [0, 1); calculate the bisector angle Angle = θ / N; The coordinates of the center point of the labeled rectangle are used as the coordinates of the equidistant points. The coordinates of the i-th equidistant point are... ,but: ; Where r is the radius, It is the starting angle, and i is from 0 to N-1.

[0010] As described above, in the edge feature extraction method for deep learning-based object recognition, the annotation includes circle annotations: Determine the coordinates of the center of the circle (h, k) and the radius r of the circle; Calculate the circumference of the circle: L = 2πr; The calculation requires generating N labeled rectangles. ; Boxsize is a pre-defined unit for the length and width of the labeled rectangle, and it sets the overlap ratio between the rectangles. , The range is [0, 1); Calculate the angle bisector Angle = 360° / N; The coordinates of the center point of the labeled rectangle are used as the coordinates of the equidistant points. The coordinates of the i-th equidistant point are... ,but: ; Where i is from 0 to N-1.

[0011] As described above, the edge feature extraction method based on deep learning for target recognition consists of two convolutional layers in the residual block of the CNN model. The first convolutional layer uses a 1×1 convolutional kernel, and the second convolutional layer uses a 3×3 convolutional kernel. The output of the residual block is the direct sum of the outputs of the two convolutional layers and the input. A 3×3 convolutional layer with a stride of 2 is used for downsampling. Loss calculation during backpropagation in model training ; in, and It is the loss for predicting the center point of the bounding box; It is the confidence loss, including the confidence loss of the foreground and background; Fixed anchor frame size: Set all anchor frames to the same value in the configuration file; The relative coordinates of the prediction box used during model training are (tx, ty), where (tx, ty) is the displacement of the center coordinates of the prediction box relative to the coordinates (Cx, Cy) of the top left corner of the grid.

[0012] As described above, the edge feature extraction method for deep learning-based object recognition includes methods for determining the inflection point of grayscale changes in the detection range, such as: The detection range of the image to be detected is converted to grayscale to obtain a grayscale image. The grayscale change rate of the grayscale image in the horizontal and vertical directions is calculated. By analyzing the local extreme points of the grayscale change rate, the inflection point of grayscale change is determined.

[0013] As described above, the edge feature extraction method based on deep learning for target recognition filters out noise at grayscale inflection points by using a threshold value of T and a direction. The T value is the difference between adjacent pixels, and the direction is the direction of the largest difference.

[0014] As described above, the edge feature extraction method based on deep learning for target recognition uses the least squares method to fit boundary feature points and feature points to form edge lines.

[0015] Compared with existing technologies, the advantages and positive effects of this invention are as follows: The edge feature extraction method for deep learning-based target recognition in this invention includes: acquiring the image to be detected and determining the detection range (blue box) of the image; determining the grayscale change inflection point of the detection range; determining the training range (red box) based on the detection range (blue box); outputting the boundary feature points of the training range (red box) through a CNN model; determining the grayscale change inflection point within a set threshold radius of the boundary feature points as feature points; and fitting the boundary feature points and feature points to form an edge line. This invention combines the innovative deep learning detection with the boundary feature point filtering function in the boundary algorithm, uses a CNN model for initial localization, improves stability in noisy environments, determines the approximate range of the boundary, and uses a boundary algorithm to locate the accurate position within the approximate range. On the one hand, it has the ability of the CNN model to maintain stable detection performance under abnormal data, noise, or environmental changes; on the other hand, it has the advantage of high precision in boundary extraction, can optimize the boundary position, and achieves sub-pixel accuracy, significantly improving the robustness and stability of detection.

[0016] Other features and advantages of the present invention will become clearer after reading the detailed embodiments of the invention in conjunction with the accompanying drawings. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of line annotations in a specific embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of the arc markings in a specific embodiment of the present invention.

[0020] Figure 3 This is a schematic diagram of a specific embodiment of the present invention with circular annotations.

[0021] Figure 4 This is a diagram of the CNN model used in a specific embodiment of the present invention.

[0022] Figure 5 This is a diagram illustrating the effect of filtering impurities in a specific embodiment of the present invention.

[0023] Figure 6 This is a diagram of the original inflection point in a specific embodiment of the present invention.

[0024] Figure 7 This is an inflection point diagram after filtering by T value in a specific embodiment of the present invention.

[0025] Figure 8 This is a schematic diagram of boundary feature points obtained by a CNN model in a specific embodiment of the present invention.

[0026] Figure 9 This is a schematic diagram of the boundary obtained by fitting a specific embodiment of the present invention.

[0027] Figure 10 This is a flowchart of boundary detection in a specific embodiment of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] In the description of this invention, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0030] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. In the description of embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0031] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0032] In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0033] An edge feature extraction method based on deep learning for object recognition includes: Obtain the image to be detected and determine the detection range of the image (blue box); Determine the inflection point of grayscale change in the detection range; determine the training range (red box) based on the detection range (blue box), and output boundary feature points for the training range using a CNN model; Determine the boundary feature points by setting the grayscale change inflection points within the threshold radius as feature points; Boundary feature points and feature points are fitted to form edge lines.

[0034] like Figure 10 As shown, the edge feature extraction method based on deep learning for object recognition includes the following steps: S1. Obtain the image to be detected.

[0035] S2. Determine the detection range of the image to be detected.

[0036] S3. Determine the grayscale inflection point of the detection range; determine the training range based on the detection range, and output boundary feature points for the training range through the CNN model.

[0037] In this step, determining the grayscale change inflection points and boundary feature points does not have a specific order and can be done simultaneously.

[0038] S4. Determine the boundary feature points. Set the gray level change inflection points within the threshold radius as feature points.

[0039] S5. Fit the boundary feature points and feature points to form the edge line.

[0040] The method for determining the training range based on the detection range is as follows: Mark the edges within the detection range and determine the number N of labeled rectangles that need to be generated within the detection range; The coordinates of the center point of the labeled rectangle are used as the coordinates of the equidistant points. The coordinates of the i-th equidistant point are... ; The coordinates of the diagonal of the training range, C(Cx, Cy) and D(Dx, Dy), are determined by the coordinates of the detection range. Cx=min(x), Cy=min(y); Dx=max(x), Dy=max(y); The labels are transformed based on the coordinates of the equidistant points and the training range, and the transformed labels are then input into the CNN model.

[0041] In traditional CNN models, annotation methods are limited to rectangles or polygons, which are not suitable for boundary annotation. Therefore, this embodiment designs line annotation, circle annotation, and arc annotation. The three annotation methods are explained below: Line annotations consist of two annotations: 1. Rotatable rectangle (red box): Used to set the detection range.

[0042] 2. Line: Used to mark the location of boundaries.

[0043] Creation method: First, identify the starting point of the straight line. Second, identify the ending point of the straight line. Third, identify the detection range.

[0044] exist Figure 1 In the examples, line annotation methods include: Determine the coordinates (Ax, Ay) and (Bx, By) of the two endpoints A and B of the line; Calculate the length of the line ; The calculation requires generating N labeled rectangles (yellow boxes). ; Boxsize is used to predetermine the length and width of the labeled rectangle, and sets the overlap ratio between the boxes. , The range is [0, 1).

[0045] The coordinates of the center point of the labeled rectangle are used as the coordinates of the equidistant points. The coordinates of the i-th equidistant point are... ,but: ; Where i ranges from 0 to N-1.

[0046] An arc annotation consists of two annotations: 1. Partial circular area: Used to set the detection range of the arc.

[0047] 2. Arc: Used to mark the boundary of an arc.

[0048] Creation method: First, identify the starting point of the arc. Second, identify any point on the arc. Third, identify the ending point of the arc. Fourth, identify the range of the arc training image.

[0049] exist Figure 2 In examples, the methods for arc labeling include: Determine the coordinates (h,k) of the center of the circle containing the arc, the central angle θ of the arc, and the radius r of the circle containing the arc; Calculate the length of the arc ; The calculation requires generating N labeled rectangles. Boxsize is the pre-defined length and width of the labeled rectangle, and sets the overlap ratio between the boxes. , The range is [0, 1); Calculate the bisector angle Angle = θ / N; The coordinates of the center point of the labeled rectangle are used as the coordinates of the equidistant points. The coordinates of the i-th equidistant point are... ,but: ; Where r is the radius, It is the starting angle, the angle of the yellow arrow in the diagram, where i is from 0 to N-1.

[0050] A circle symbol consists of two symbols: 1. Circular area: Used to set the detection range of the circle.

[0051] 2. Circle: Used to mark the boundary of a circle.

[0052] Creation method: The first three points confirm the boundary of the circle. The fourth point confirms the detection range of the circle.

[0053] exist Figure 3 In the examples, the methods for circle annotation include: Determine the coordinates of the center of the circle (h, k) and the radius r of the circle; Calculate the circumference of the circle: L = 2πr; The calculation requires generating N labeled rectangles. ; Boxsize is where the length and width of the labeled rectangle are predetermined, and the overlap ratio between the rectangles is set. , The range is [0, 1), which sets the overlap ratio between boxes; Calculate the bisector angle Angle = θ / N; The coordinates of the center point of the labeled rectangle are used as the coordinates of the equidistant points. The coordinates of the i-th equidistant point are... ,but: ; Where i ranges from 0 to N-1. Because standard CNN models do not support edge detection annotations, the above annotations need to be converted into universal bounding boxes supported by CNN models. Before outputting boundary feature points through the CNN model, the annotations of the detection range need to be transformed. Methods for annotation transformation include: Determine the number N of labeled rectangles that need to be generated within the detection range; The coordinates of the center point of the labeled rectangle are used as the coordinates of the equidistant points. The coordinates of the i-th equidistant point are... ; The coordinates of the diagonal of the training range, C(Cx, Cy) and D(Dx, Dy), are determined by the coordinates of the detection range. Cx=min(x), Cy=min(y)); Dx=max(x), Dy=max(y)); Annotation generation: <object-class>: The category ID of the target object, numbered starting from 0; <x_center> =(Pi(x)-Cx) / width w,<x_center> The x-coordinate of the center point of the labeled rectangle is normalized to the range of 0 to 1 based on the image width. <y_center> =(Pi(y)-Cy) / heighth,<y_center> The y-coordinate of the center point of the labeled rectangle is normalized to the image height and ranges from 0 to 1. <width>=Boxsize / width w, <width>To define the width of the bounding box, normalize it to the image width to a range between 0 and 1; <height>=Boxsize / height h, <height>To annotate the height of the rectangular frame, the range normalized to the image height is between 0 and 1; where width w = max x - min x; height h = max y - min y.

[0054] The network composition of the CNN model mainly consists of convolutional layers, residual blocks (ResidualBlock), and pooling layers.

[0055] The network is divided into the front section, the middle section, and the back section: The front section is mainly composed of convolutional layers and max pooling layers, which are used to extract the initial features of the image.

[0056] The middle section is mainly composed of residual blocks, which enhance the feature extraction ability and accelerate the convergence through residual connections.

[0057] The back section outputs the final prediction results through the global average pooling layer and the fully connected layer in the classification task.

[0058] Residual block structure: Each residual block of the CNN model consists of two convolutional layers: The first convolutional layer uses a 1x1 convolutional kernel to reduce the number of channels.

[0059] The second convolutional layer uses a 3x3 convolutional kernel for feature extraction.

[0060] The output of the residual block is the direct addition of the outputs of the two convolutional layers and the input, which prevents gradient vanishing.

[0061] Downsampling strategy: using 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.

[0062] Output features: output three different size feature maps (13x13, 26x26, 52x52) for multi-scale target detection.

[0063] 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.

[0064] Loss calculation of backpropagation ; where, and are the losses of the center points of the predicted boxes; is the confidence loss, including the confidence loss of foreground and background; The relative coordinate of the prediction frame in the model training process is (tx, ty), which is the displacement of the center coordinate of the prediction frame relative to the top-left corner coordinate (Cx, Cy) of the grid.

[0065] The CNN model is trained and saved.

[0066] The method for determining the gray scale change inflection point of the detection range comprises: 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.

[0067] 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:

[0068] The gray scale change inflection points are determined by analyzing the local extreme points of the gray scale change rate. The gray scale change inflection points are points on the gray scale change rate curve, and the calculation thereof is obtained by first-order derivation. The calculation formula is as follows: ; For the gray scale change inflection points, the T value, which is the difference between adjacent pixels, and the direction, which is the maximum difference direction, are used to filter the noise points.

[0069] The T value, which is the difference between adjacent pixels, and the direction, which is the maximum difference direction, are used to filter the noise points, and the filtering effect is shown in Figure 5 .

[0070] The T value is the difference between adjacent pixels, and the direction is the maximum difference direction. The threshold refers to the difference within a certain range, for example, the value between 10 and 50.

[0071] For example, the original inflection point graph is shown in Figure 6 , and the inflection point graph after T value filtering is shown in Figure 7 .

[0072] The boundary feature points of the training range are detected by the above-mentioned CNN model, and the result is shown in Figure 8 .

[0073] The gray scale change inflection points within the threshold radius range of the boundary feature points are determined as the feature points.

[0074] The boundary feature points and the feature points are fitted to form an edge line.

[0075] The boundary feature points and the feature points are fitted to form an edge line by using the least squares method, and the final boundary effect is shown in Figure 9 .

[0076] The resulting curved edges can better reflect the true contours of objects in the image, improving the effectiveness of edge recognition.

[0077] This implementation example has the following beneficial effects: 1. Improve annotation efficiency.

[0078] 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.

[0079] 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.

[0080] Improved annotation accuracy: The new annotation format reduces human error due to fewer operations, thereby improving the overall accuracy of annotations.

[0081] Therefore, the innovative deep learning annotation format not only significantly improves annotation efficiency but also enables automatic annotation.

[0082] 2. Improve detection accuracy and the ability to capture complex boundaries: 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.

[0083] 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.

[0084] 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.

[0085] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions claimed by the present invention.< / height> < / height> < / width> < / width>

Claims

1. An edge feature extraction method based on deep learning for target recognition, characterized in that, The method includes: Obtain the image to be detected and determine the detection range of the image to be detected; Determine the grayscale inflection point of the detection range; determine the training range based on the detection range, and output boundary feature points for the training range using a CNN model; The grayscale inflection points within a set threshold radius of the boundary feature points are identified as feature points. The boundary feature points and feature points are fitted to form the edge line.

2. The edge feature extraction method based on deep learning target recognition according to claim 1, characterized in that, The method for determining the training range based on the detection range is as follows: The edges within the detection range are marked, and the number N of marked rectangles that need to be generated within the detection range is determined. The coordinates of the center point of the labeled rectangle are used as the coordinates of the equidistant points. The coordinates of the i-th equidistant point are... ; The coordinates of the diagonal of the training range, C(Cx, Cy) and D(Dx, Dy), are determined by the coordinates of the detection range. Cx=min(x), Cy=min(y)); Dx=max(x), Dy=max(y)); The labels are transformed based on the coordinates of the equidistant points and the training range, and the transformed labels are input into the CNN model.

3. The edge feature extraction method based on deep learning target recognition according to claim 2, characterized in that, The annotations include line annotations: Determine the coordinates (Ax, Ay) and (Bx, By) of the two endpoints A and B of the line; Calculate the length of the line ; The calculation requires generating N labeled rectangles. ; Boxsize is used to predetermine the length and width of the labeled rectangle, and sets the overlap ratio between the boxes. , The range is [0, 1); The coordinates of the center point of the labeled rectangle are used as the coordinates of the equidistant points. The coordinates of the i-th equidistant point are... ,but: ； Where i ranges from 0 to N-1.

4. The edge feature extraction method based on deep learning target recognition according to claim 2, characterized in that, The annotations include arc annotations: Determine the coordinates (h,k) of the center of the circle containing the arc, the central angle θ of the arc, and the radius r of the circle containing the arc; Calculate the length of the arc ; The calculation requires generating N labeled rectangles. ; Boxsize is a pre-defined unit for the length and width of the labeled rectangle, and it sets the overlap ratio between the rectangles. , The range is [0, 1); calculate the bisector angle Angle = θ / N; The coordinates of the center point of the labeled rectangle are used as the coordinates of the equidistant points. The coordinates of the i-th equidistant point are... ,but: ； Where r is the radius, It is the starting angle, and i is from 0 to N-1.

5. The edge feature extraction method based on deep learning target recognition according to claim 2, characterized in that, The annotations include circle annotations: Determine the coordinates of the center of the circle (h, k) and the radius r of the circle; Calculate the circumference of the circle: L = 2πr; The calculation requires generating N labeled rectangles. ; Boxsize is a pre-defined unit for the length and width of the labeled rectangle, and it sets the overlap ratio between the rectangles. , The range is [0, 1); Calculate the angle bisector Angle = 360° / N; The coordinates of the center point of the labeled rectangle are used as the coordinates of the equidistant points. The coordinates of the i-th equidistant point are... ,but: ； Where i is from 0 to N-1.

6. 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 consists of two convolutional layers. The first convolutional layer uses a 1×1 convolutional kernel, and the second convolutional layer uses a 3×3 convolutional kernel. The output of the residual block is the direct sum of the outputs of the two convolutional layers and the input. A 3×3 convolutional layer with a stride of 2 is used for downsampling. Loss calculation during backpropagation in model training ; in, and It is the loss for predicting the center point of the bounding box; It is the confidence loss, including the confidence loss of the foreground and background; Fixed anchor frame size: Set all anchor frames to the same value in the configuration file; The relative coordinates of the prediction box used during model training are (tx, ty), where (tx, ty) is the displacement of the center coordinates of the prediction box relative to the coordinates (Cx, Cy) of the top left corner of the grid.

7. The edge feature extraction method based on deep learning target recognition according to claim 1, characterized in that, Methods for determining the inflection point of grayscale change in the detection range include: The detection range of the image to be detected is converted to grayscale to obtain a grayscale image. The grayscale change rate of the grayscale image in the horizontal and vertical directions is calculated. By analyzing the local extreme points of the grayscale change rate, the inflection point of grayscale change is determined.

8. The edge feature extraction method based on deep learning target recognition according to claim 7, characterized in that, For grayscale inflection points, noise is filtered out by the magnitude threshold and direction of the T value. The T value is the difference between adjacent pixels, and the direction is the direction of the greatest difference.

9. The edge feature extraction method based on deep learning target recognition according to claim 1, characterized in that, The boundary feature points and feature points are fitted using the least squares method to form the edge line.

Images