Siamese network target tracking method based on hybrid structure attention guidance

Abstract

The present invention is based on a twin network target tracking method guided by hybrid structure attention, specifically: dividing the target tracking data set into a training set and a test set, and then preprocessing the data set; constructing a hybrid structure attention feature fusion module; constructing a twin network target tracking algorithm model based on hybrid structure attention guidance; using the training set to iteratively train the twin network target tracking algorithm model based on hybrid structure attention guidance constructed in step 3 to obtain a trained twin network target tracking model based on a hybrid structure attention mechanism; using the test set as the input of the trained twin network target tracking model based on a hybrid structure attention mechanism for forward reasoning to obtain the tracking result of each test sample. The method of the present invention solves the problem of low precision of existing target tracking methods.

Description

Technical Field

[0001] The present invention belongs to the technical field of target tracking methods, and specifically relates to a twin network target tracking method based on hybrid structure attention guidance. Background Art

[0002] Target tracking is a technology widely used in fields such as robotic vision, video surveillance, missile guidance, and unmanned driving. To achieve high-precision and robust target tracking, effective target tracking algorithms are necessary. Tracking targets in real-world scenarios presents numerous challenges, including target deformation, occlusion, and interference from similar targets. To better address these challenges, twin network target tracking algorithms have emerged. However, current twin network target tracking algorithms simply fuse template features and search region features through cross-correlation, losing rich semantic information. To address this issue, a Transformer-based twin network target tracking algorithm was proposed. Due to its advantages in modeling long-range dependencies, the Transformer is used in the feature fusion stage of twin network target tracking to integrate information between the template and the search region. However, the Transformer lacks the advantage of learning local details, making the tracking algorithm susceptible to interference from similar targets, leading to suboptimal solutions and becoming a bottleneck in the design of high-precision target tracking algorithms. Summary of the Invention

[0003] The purpose of this invention is to provide a twin network target tracking method based on hybrid structure attention guidance, which solves the problem of low accuracy of existing target tracking methods.

[0004] The technical solution adopted by the present invention is a twin network target tracking method based on hybrid structure attention guidance, which is specifically implemented according to the following steps:

[0005] Step 1: Divide the target tracking dataset into a training set and a test set, and then preprocess the dataset;

[0006] Step 2: Construct a hybrid structure attention feature fusion module;

[0007] Step 3: Build a twin network target tracking algorithm model based on hybrid structure attention guidance;

[0008] Step 4: Use the training set to iteratively train the twin network target tracking algorithm model based on hybrid structure attention guidance constructed in step 3 to obtain a trained twin network target tracking model based on hybrid structure attention mechanism;

[0009] In step 5, the test set is used as the input of the trained twin network target tracking model based on the hybrid structure attention mechanism for forward reasoning to obtain the tracking results of each test sample.

[0010] The present invention is also characterized in that:

[0011] In step 1, the specific process of preprocessing the target tracking dataset is as follows:

[0012] S1: capture the target area of the initial video frame as the calibration template, and scale the calibration template to a size of 128×128;

[0013] S2: In the current frame, an image area with a search size of 256×256 is captured with the target as the center as the search area image;

[0014] S3: Capture the target area of the current frame as a dynamic template, and scale the dynamic template to a size of 128×128.

[0015] The specific process of step 2 is:

[0016] Suppose image feature x m and image features x n , the image feature x m and image features x n The convolution layer composed of convolution kernel size of 3×3 learns local detail information and introduces position information to obtain the image feature x' m and image features x' n , let image feature x' m The query vector x obtained by the linear layer q , image feature x' n The value vector x obtained by the linear layer v and key vector x k , after the multi-head mutual attention layer in the hybrid structure attention, the image features x' that can be reflected in different channel dimensions are obtained m and image features x' n The attention map of the degree of correlation between them is expressed as:

[0017]

[0018] Where x s is the image feature x' m Attention map, d represents the dimension of the key vector;

[0019] The attention map x output by the multi-head mutual attention layer s With image feature x' m Added through the residual unit, the image feature x' is realized m and image features x' n Aggregation of effective information.

[0020] In step 3, the twin network target tracking algorithm model based on hybrid structure attention guidance includes three parts: feature extraction network, feature fusion network, and classification and regression network.

[0021] The specific process of step 3 is:

[0022] Step 3.1: The search area image, calibration template image, and dynamic template image obtained in step 1 are respectively passed through the feature extraction network ResNet50 to obtain calibration template features, dynamic template features, and search area features;

[0023] Step 3.2: Use the feature fusion network to fuse the calibration template features and dynamic template features into template features. Then fuse the template features with the search area features to obtain the attention map of the search area features and the attention map of the template features. Add the search area features and the attention map of the search area features and pass them through the residual unit to output the search area features that integrate the target position information. Add the template features and the attention map of the template features and pass them through the residual unit to output the template features that integrate the target key information.

[0024] In step 3.3, the classification and regression network is used to parse the search area features that integrate the target location information and the template features that integrate the target key information output in step 3.2 to obtain the tracking target frame.

[0025] In step 3.1, the feature extraction network ResNet50 is a twin network structure. The stride parameter of the third layer in ResNet50 is set to 1, so that the output feature map of the third layer and the second layer are the same size, and the input image features are obtained from the third layer.

[0026] The specific process of step 3.2 is as follows: the information of the calibration template features and dynamic template features obtained in step 3.1 is aggregated using the hybrid structure attention feature fusion module constructed in step 2 to obtain the template features, and then the template features are fused with the search area features obtained in step 3.1 using the symmetrical hybrid structure attention module;

[0027] The specific process of fusing the template features and the search area features obtained in step 3.1 using the symmetrical hybrid structure attention module is as follows:

[0028] The search area features obtained in step 3.1 are enhanced by the convolution layer consisting of convolution kernels of 1×1 and 3×3. The template features obtained in step 3.1 are enhanced by the convolution layer consisting of convolution kernels of 1×1 and 3×3. Suppose the search area features are obtained by the linear layer to obtain the query vector s q , value vector s v and key vector sk , the template features are passed through the linear layer to obtain the query vector t q , value vector t v and key vector t k , after the multi-head mutual attention layer in the hybrid structure attention, the attention maps that can reflect the degree of correlation between the search area features and the template features in different channel dimensions are obtained respectively. The expression is:

[0029]

[0030]

[0031] Where a s is the attention map of the search area feature, which contains the location information of the target, a t The attention map representing the template features contains key information that can identify the target, and d represents the dimension of the key vector;

[0032] Search region features and attention map a s After addition, the residual unit is used to output the search area features that integrate the target location information;

[0033] Template features and attention mapa t After addition, the residual unit is used to output the template features that integrate the key information of the target.

[0034] In step 3.3, the structures of the classification and regression networks are as follows: the classification network consists of three convolutions with a kernel size of 3, a stride of 1, and a padding size of 1; the regression network consists of three convolutions with a kernel size of 3, a stride of 1, and a padding size of 1, and one convolution with a kernel size of 5, a stride of 1, and a padding size of 2.

[0035] In step 4, the loss function is used for back propagation during training to update the model weight parameters. The loss function is as follows:

[0036] L=λ1L reg +λ2L cls

[0037] Where, L reg is the loss value of the regression task, L cls is the loss value of the classification task, λ1 is the weight of the regression loss, and λ2 is the weight of the classification loss;

[0038] in,

[0039]

[0040] Where IOU represents the ratio of the intersection and union of the label box area and the predicted tracking target box area, y regIndicates the offset value of the upper left corner coordinate and the lower right corner coordinate of the label box to the reference coordinate, y reg =(x l ,y l ,x r ,y r ), x l and y l Is the offset value of the upper left corner coordinate of the label box relative to the reference coordinate, x r and y r It is the offset value of the lower right corner coordinate of the label box relative to the reference coordinate. Indicates the offset value of the upper left corner coordinate and the lower right corner coordinate of the predicted tracking target box to the reference coordinate. and It is the offset value of the upper left corner coordinate of the predicted tracking target frame relative to the reference coordinate. and It is the offset value of the lower right corner coordinate of the predicted tracking target box relative to the reference coordinate;

[0041] in,

[0042]

[0043] Where M represents the number of positive samples and negative samples, is the label value of the i-th sample, y i is the predicted value of the i-th sample.

[0044] Positive samples and negative samples are divided by dividing the elliptical area in the label box. The center point of the label box is used as the center of the ellipse, the length of the label box is used as the major axis of the ellipse, and the width is used as the minor axis. If the sampling point falls outside the elliptical area, it is a negative sample. The center point remains unchanged, half of the length of the label box is used as the major axis of the ellipse, and half of the width is used as the minor axis. A small elliptical area is divided, and if it falls in this area, it is a positive sample.

[0045] The beneficial effects of the present invention are: the method of the present invention combines convolution and Transformer to realize a feature fusion module based on hybrid structure attention, which can perform local fine-grained feature learning and global feature aggregation, and constructs a twin network target tracking algorithm model based on hybrid structure attention guidance, realizing information integration of template features and search area features, and improving the tracking algorithm's ability to accurately locate targets. At the same time, a dynamic template is added to the input end of the template branch of the twin network target tracking algorithm model based on hybrid structure attention guidance, that is, the calibration template and the dynamic template are input at the same time, and the hybrid structure attention is used to learn and concentrate the fine-grained features of the target in the calibration template and the dynamic template to obtain template features that can adapt to changes in target morphology. The method of the present invention can robustly track targets in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a flow chart of the target tracking method of the twin network based on hybrid structure attention guidance of the present invention;

[0047] Figure 2 Flowchart of data set preprocessing in the method of the present invention;

[0048] Figure 3 A comparison chart of the success rates of the method of the present invention and the prior art;

[0049] Figure 4 This is a comparison chart of the accuracy of the method of the present invention and the prior art. DETAILED DESCRIPTION

[0050] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] The present invention is based on a twin network target tracking method guided by hybrid structure attention, such as Figure 1 As shown, please follow the steps below:

[0052] Step 1: According to the Pareto principle, 80% of the image data in the target tracking dataset is used as the training set, and 20% of the images are used as the test set. The training set includes image data and target label box information, and the test set only contains image data. The dataset is then preprocessed to obtain the search area image, calibration template image, and dynamic template image.

[0053] like Figure 2 As shown in Figure 2, the specific process of preprocessing the target tracking dataset is as follows:

[0054] S1: capture the target area of the initial video frame as the calibration template, and scale the calibration template to a size of 128×128;

[0055] S2: In the current frame, an image area with a search size of 256×256 is captured with the target as the center as the search area image;

[0056] S3: intercept the target area of the current frame as a dynamic template, and scale the dynamic template to a size of 128×128;

[0057] Step 2: Construct a hybrid structure attention feature fusion module;

[0058] The specific process is:

[0059] Suppose image feature x m and image features x n , the image feature x m and image features x nThe convolution layer composed of convolution kernel size of 3×3 learns local detail information and introduces position information to obtain the image feature x' m and image features x' n , let image feature x' m The query vector x obtained by the linear layer q , image feature x' n The value vector x obtained by the linear layer v and key vector x k , after the multi-head mutual attention layer in the hybrid structure attention, the image features x' that can be reflected in different channel dimensions are obtained m and image features x' n The attention map of the degree of correlation between them is expressed as:

[0060]

[0061] Where x s is the image feature x' m Attention map, d represents the dimension of the key vector;

[0062] The attention map x output by the multi-head mutual attention layer s With image feature x' m Added through the residual unit, the image feature x' is realized m and image features x' n Aggregation of effective information in

[0063] Step 3: Construct a twin network target tracking algorithm model based on hybrid structure attention guidance. The twin network target tracking algorithm model based on hybrid structure attention guidance includes three parts: feature extraction network, feature fusion network, and classification and regression network;

[0064] Step 3.1: The search area image, calibration template image, and dynamic template image obtained in step 1 are respectively passed through the feature extraction network ResNet50 to obtain calibration template features, dynamic template features, and search area features;

[0065] The feature extraction network is a twin network structure, specifically using ResNet50 and setting the stride parameter of the third layer in ResNet50 to 1, so that the output feature map of the third layer is the same size as the second layer, and the input image features are obtained from the third layer;

[0066] Step 3.2: Use the feature fusion network to fuse the calibration template features and dynamic template features into template features. Then fuse the template features with the search area features to obtain the attention map of the search area features and the attention map of the template features. Add the search area features and the attention map of the search area features and pass them through the residual unit to output the search area features that integrate the target position information. Add the template features and the attention map of the template features and pass them through the residual unit to output the template features that integrate the target key information.

[0067] The specific process is:

[0068] The hybrid structure attention feature fusion module constructed in step 2 is used to aggregate the information of the calibration template features and dynamic template features obtained in step 3.1 to obtain the template features, which can accurately represent the target appearance model. The template features are then fused with the search area features obtained in step 3.1 using a symmetrical hybrid structure attention module.

[0069] The specific process of fusing the template features and the search area features obtained in step 3.1 using the symmetrical hybrid structure attention module is as follows:

[0070] The search area features obtained in step 3.1 are enhanced by the convolution layer consisting of convolution kernels of 1×1 and 3×3. The template features obtained in step 3.1 are enhanced by the convolution layer consisting of convolution kernels of 1×1 and 3×3. Suppose the search area features are obtained by the linear layer to obtain the query vector s q , value vector s v and key vector s k , the template features are passed through the linear layer to obtain the query vector t q , value vector t v and key vector t k , after the multi-head mutual attention layer in the hybrid structure attention, the attention maps that can reflect the degree of correlation between the search area features and the template features in different channel dimensions are obtained respectively. The expression is:

[0071]

[0072]

[0073] Where a s is the attention map of the search area feature, which contains the location information of the target, a t The attention map representing the template features contains key information that can identify the target, and d represents the dimension of the key vector;

[0074] Search region features and attention map a sAfter addition, the residual unit is used to output the search area features that integrate the target location information;

[0075] Template features and attention mapa t After addition, the residual unit is used to output the template features that integrate the key information of the target;

[0076] In this process, the search area features and template features can be fused multiple times to gradually focus attention on the target. In this invention, the fusion is performed four times. Finally, the feature fusion network outputs the search area features containing the target location information.

[0077] Step 3.3: Use the classification and regression network to parse the search area features that integrate the target location information and the template features that integrate the target key information output in step 3.2 to obtain the tracking target frame;

[0078] The structure of the classification and regression networks is as follows: the classification network consists of three convolutions with a kernel size of 3, a stride of 1, and a padding size of 1; the regression network consists of three convolutions with a kernel size of 3, a stride of 1, and a padding size of 1, and one convolution with a kernel size of 5, a stride of 1, and a padding size of 2;

[0079] Step 4: Use the training set to iteratively train the twin network target tracking algorithm model based on hybrid structure attention guidance constructed in step 3. During the training process, the loss function is used for back propagation to update the model weight parameters to obtain the trained twin network target tracking model based on hybrid structure attention mechanism;

[0080] The present invention adopts a twin network target tracking algorithm model based on hybrid structure attention guidance to jointly calculate the classification task loss and regression task loss. The loss function is as follows:

[0081] L=λ1L reg +λ2L cls

[0082] Where, L reg is the loss value of the regression task, L cls is the loss value of the classification task, λ1 is the weight of the regression loss, λ1=1.0, λ2 is the weight of the classification loss, λ2=1.0;

[0083] The weight parameters λ1 and λ2 are introduced when calculating the loss. By adjusting the weight parameters, the model's ability to distinguish between targets and backgrounds is enhanced, improving the tracker's tracking ability in complex scenarios.

[0084] The hybrid structure attention-guided twin network target tracking algorithm model constructed in this paper adopts an anchor-free frame structure, so the size of the tracking target frame can be directly predicted. The loss value of the regression task is obtained by calculating the ratio of the intersection and union of the label frame area and the predicted tracking target frame area. The specific expression is as follows:

[0085]

[0086] Where IOU represents the ratio of the intersection and union of the label box area and the predicted tracking target box area. When the IOU value is 1, the label box area and the tracking box area completely overlap, and the loss value is 0. reg Indicates the offset value of the upper left corner coordinate and the lower right corner coordinate of the label box to the reference coordinate, y reg =(x l ,y l ,x r ,y r ), x l and y l Is the offset value of the upper left corner coordinate of the label box relative to the reference coordinate, x r and y r It is the offset value of the lower right corner coordinate of the label box relative to the reference coordinate. Indicates the offset value of the upper left corner coordinate and the lower right corner coordinate of the predicted tracking target box to the reference coordinate. and It is the offset value of the upper left corner coordinate of the predicted tracking target frame relative to the reference coordinate. and It is the offset value of the lower right corner coordinate of the predicted tracking target frame relative to the reference coordinate. The IOU loss function is used in the regression task to optimize the target frame as a whole, thereby improving the quality of the predicted tracking target frame.

[0087] The loss value of the classification task adopts the binary cross entropy loss function, and the specific expression is as follows:

[0088]

[0089] Where M represents the number of positive samples and negative samples, is the label value of the i-th sample, y i is the predicted value of the i-th sample;

[0090] Positive samples and negative samples are divided by dividing the elliptical area in the label box. The center point of the label box is used as the center of the ellipse, the length of the label box is used as the major axis of the ellipse, and the width is used as the minor axis. If the sampling point falls outside the elliptical area, it is a negative sample. The center point remains unchanged, half of the length of the label box is used as the major axis of the ellipse, and half of the width is used as the minor axis. A small elliptical area is divided, and if it falls in this area, it is a positive sample.

[0091] In step 5, the test set is used as the input of the trained twin network target tracking model based on the hybrid structure attention mechanism for forward reasoning to obtain the tracking results of each test sample.

[0092] The experimental comparison results of the method of the present invention and the prior art are as follows:

[0093] The OTB100 dataset, proposed by Wu et al. and widely used in object tracking, contains 100 challenging test scenes composed of both grayscale and color data. These scenes are subject to common challenges in visual tracking, such as scale variation, occlusion, deformation, background interference, rotation, and motion blur. The OTB100 dataset evaluates the performance of object tracking algorithms based on tracking accuracy and tracking success rate. Tracking accuracy is the error between the center position of the tracking result and the true target center, displayed in the Precision Plot; tracking success rate reflects the overlap between the tracking box and the true target box, displayed in the Success Plot.

[0094] In this dataset, this method is compared with several current excellent methods, such as Figure 3 As shown in the figure, the target tracking methods are ranked according to the Area Under Curve (AUC) score. Compared with the current excellent methods, the proposed method has the best AUC score and Precision score. The AUC score exceeds the second place by 1.2%, which proves that enhancing the model's ability to learn detailed information can reduce tracking drift and improve the tracking success rate. Figure 4 As shown in the figure, the target tracking methods are ranked according to the precision score (Precision) with a center distance error threshold (Location error threshold) of 20. The precision score of the proposed method exceeds the second place by 0.6%. Combining the tracking success rate and tracking accuracy results, it is proved that the proposed method can achieve excellent tracking performance.

Claims

1. A twin network target tracking method based on hybrid structure attention guidance, characterized by: Please follow the steps below to implement it: Step 1: Divide the target tracking dataset into a training set and a test set, and then preprocess the dataset; Step 2: Construct a hybrid structure attention feature fusion module; The specific process of step 2 is: With image features and image features , the image features and image features After the convolution kernel size is 3 The convolution layer composed of 3 convolutions learns local detail information and introduces position information to obtain image features and image features , let the image feature The query vector obtained by the linear layer , image features The value vector obtained by the linear layer and key vector , through the multi-head mutual attention layer in the hybrid structure attention, the image features can be reflected in different channel dimensions and image features The attention map of the degree of correlation between them is expressed as: Where, is the image feature Attention map, represents the dimension of the key vector; The attention map output by the multi-head mutual attention layer and image features Adding through the residual unit, the image features are realized and image features Aggregation of effective information in Step 3: Build a twin network target tracking algorithm model based on hybrid structure attention guidance; In step 3, the twin network target tracking algorithm model based on hybrid structure attention guidance includes three parts: feature extraction network, feature fusion network, and classification and regression network; Step 4: Use the training set to iteratively train the twin network target tracking algorithm model based on hybrid structure attention guidance constructed in step 3 to obtain a trained twin network target tracking model based on hybrid structure attention mechanism; In step 5, the test set is used as the input of the trained twin network target tracking model based on the hybrid structure attention mechanism for forward reasoning to obtain the tracking results of each test sample.

2. The target tracking method based on a hybrid structure attention guidance twin network according to claim 1 is characterized in that: In step 1, the specific process of preprocessing the target tracking dataset is as follows: S1: capture the target area of the initial video frame as the calibration template, and scale the calibration template to a size of 128×128; S2: In the current frame, an image area with a search size of 256×256 is captured with the target as the center as the search area image; S3: Capture the target area of the current frame as a dynamic template, and scale the dynamic template to a size of 128×128.

3. The target tracking method based on a hybrid structure attention guidance twin network according to claim 1 is characterized in that: The specific process of step 3 is: Step 3.1: The search area image, calibration template image, and dynamic template image obtained in step 1 are respectively passed through the feature extraction network ResNet50 to obtain calibration template features, dynamic template features, and search area features; Step 3.2: Use the feature fusion network to fuse the calibration template features and dynamic template features into template features. Then fuse the template features with the search area features to obtain the attention map of the search area features and the attention map of the template features. Add the search area features and the attention map of the search area features and pass them through the residual unit to output the search area features that integrate the target position information. Add the template features and the attention map of the template features and pass them through the residual unit to output the template features that integrate the target key information. In step 3.3, the classification and regression network is used to parse the search area features that integrate the target location information and the template features that integrate the target key information output in step 3.2 to obtain the tracking target frame.

4. The target tracking method of a Siamese network based on hybrid structure attention guidance according to claim 3 is characterized in that: In step 3.1, the feature extraction network ResNet50 is a twin network structure. The stride parameter of the third layer in ResNet50 is set to 1, so that the output feature map of the third layer and the second layer are the same size, and the input image features are obtained from the third layer.

5. The target tracking method of a Siamese network based on hybrid structure attention guidance according to claim 3 is characterized in that: The specific process of step 3.2 is as follows: the information of the calibration template features and dynamic template features obtained in step 3.1 is aggregated using the hybrid structure attention feature fusion module constructed in step 2 to obtain the template features, and then the template features are fused with the search area features obtained in step 3.1 using the symmetrical hybrid structure attention module; The specific process of fusing the template features and the search area features obtained in step 3.1 using the symmetrical hybrid structure attention module is as follows: The search area features obtained in step 3.1 are enhanced by the convolution layer consisting of convolution kernels of 1×1 and 3×3. The template features obtained in step 3.1 are enhanced by the convolution layer consisting of convolution kernels of 1×1 and 3×3. The query vector is obtained by the linear layer. , value vector and key vector , the template features are passed through the linear layer to obtain the query vector , value vector and key vector , after the multi-head mutual attention layer in the hybrid structure attention, the attention maps that can reflect the degree of correlation between the search area features and the template features in different channel dimensions are obtained respectively. The expression is: Where, is the attention map of the search area features, which contains the location information of the target. The attention map representing the template features contains key information that can identify the target. represents the dimension of the key vector; Search region features and attention maps After addition, the residual unit is used to output the search area features that integrate the target location information; Template features and attention maps After addition, the residual unit is used to output the template features that integrate the key information of the target.

6. The Siamese network target tracking method based on hybrid structure attention guidance according to claim 3 is characterized in that In step 3.3, the structures of the classification and regression networks are as follows: the classification network consists of three convolutions with a kernel size of 3, a stride of 1, and a padding size of 1; the regression network consists of three convolutions with a kernel size of 3, a stride of 1, and a padding size of 1, and one convolution with a kernel size of 5, a stride of 1, and a padding size of 2.

7. The target tracking method based on a hybrid structure attention guidance twin network according to claim 1 is characterized in that: In step 4, the loss function is used for back propagation during training to update the model weight parameters. The loss function is as follows: Where, is the loss value of the regression task, is the loss value of the classification task, is the weight of the regression loss, is the weight of the classification loss; in, Where, It represents the ratio of the intersection and union of the label box area and the predicted tracking target box area. Indicates the offset value of the upper left corner coordinate and the lower right corner coordinate of the label box to the reference coordinate. , and It is the offset value of the upper left corner coordinate of the label box relative to the reference coordinate. and It is the offset value of the lower right corner coordinate of the label box relative to the reference coordinate. Indicates the offset value of the upper left corner coordinate and the lower right corner coordinate of the predicted tracking target box to the reference coordinate. , and It is the offset value of the upper left corner coordinate of the predicted tracking target frame relative to the reference coordinate. and It is the offset value of the lower right corner coordinate of the predicted tracking target box relative to the reference coordinate; in, Where, represents the number of positive samples and negative samples, It is The label value of each sample, It is Sample prediction values.

8. The Siamese network target tracking method based on hybrid structure attention guidance according to claim 7 is characterized in that: The positive samples and negative samples are divided by dividing the elliptical area in the label frame, with the center point of the label frame as the center of the ellipse, the length of the label frame as the major axis of the ellipse, and the width as the minor axis. If the sampling point falls outside the elliptical area, it is a negative sample. The center point remains unchanged, half of the length of the label frame is used as the major axis of the ellipse, and half of the width is used as the minor axis. A small elliptical area is divided, and the sample falling in this area is a positive sample.

Images