Remote sensing image aircraft target detection method fusing channel attention

Abstract

The invention is suitable for the technical field of spaceflight, and provides a remote sensing image aircraft target detection method fusing channel attention, and the method comprises the steps: obtaining a to-be-detected satellite remote sensing image, obtaining the target detection result of each block image, and obtaining the target detection result of the satellite remote sensing image. Through multi-scale feature extraction and feature fusion, image information under different scales can be fully utilized, and the target detection precision is improved. According to the method, a large-scale satellite remote sensing image is segmented into a plurality of small-scale block images for processing by adopting a block processing mode, so that the calculation complexity is reduced, and the detection efficiency is improved. Due to the introduction of the self-attention mechanism, the model can pay more attention to key areas in the image, and the robustness of the model to complex backgrounds and shielding conditions is improved.

Description

Technical Field

[0001] The invention belongs to the field of aerospace technology, and in particular relates to a remote sensing image aircraft target detection method integrating channel attention. Background Art

[0002] With the rapid development of aerospace technology, the image data acquired by remote sensing satellites is increasing, and the image resolution is constantly improving. However, traditional remote sensing image target detection methods have problems such as difficulty in feature extraction, low detection accuracy, and high computational complexity, which are difficult to meet the needs of practical applications. Therefore, it is particularly important to develop an efficient and accurate remote sensing satellite image target detection method.

[0003] In the prior art, remote sensing satellite image target detection usually includes steps such as image preprocessing, feature extraction, and target classification. Among them, feature extraction is the key link, but existing methods often have difficulty in effectively extracting complex features in images, especially when the target scale varies greatly and the background is complex. In addition, although some machine learning-based detection methods have achieved certain results, the training process is complicated and the computing resources are consumed a lot, which limits their promotion in practical applications. Summary of the invention

[0004] The purpose of the present invention is to provide a remote sensing satellite image target detection method based on deep learning, which can realize fast and accurate detection of targets in remote sensing satellite images by constructing an efficient target detection model, improve detection efficiency and accuracy, and reduce computational complexity.

[0005] The present invention is implemented as follows: a remote sensing image aircraft target detection method integrating channel attention, the method comprising:

[0006] S101: Obtain a satellite remote sensing image to be detected through a public data set;

[0007] S102: Divide the satellite remote sensing image into blocks, and input each block image into a target detection model to obtain a target detection result of each block image output by the target detection model;

[0008] S103: Fusing the target detection results of each block image to obtain the target detection result of the satellite remote sensing image.

[0009] As a further solution of the present invention, the target detection model is used to perform multi-scale feature extraction on each block image to obtain a multi-scale initial feature map, perform feature fusion on the multi-scale feature map to obtain a multi-scale fused feature map, perform self-attention extraction on the multi-scale fused feature map to obtain a multi-scale enhanced feature map, and perform target detection based on the multi-scale enhanced feature map to obtain a target detection result for each block image;

[0010] As a further solution of the present invention, in S102, it also includes:

[0011] Apply preprocessing strategy to the image;

[0012] Obtain positive and negative samples of any sample block image of the sample satellite remote sensing image;

[0013] For each scale in the multi-scale, a channel attention operation is performed on the initial feature map to obtain a channel attention map, which is then fused with the initial feature map to obtain a channel attention weighted map.

[0014] As a further solution of the present invention, the positive and negative samples of any sample block image of the sample satellite remote sensing image are obtained, specifically:

[0015] Use the clustering algorithm to cluster the real frames of the training set and generate multiple anchor frames;

[0016] Calculate the center distance between each real box and all anchor boxes, and select a preset number of anchor boxes with the smallest center distance as candidate anchor boxes

[0017] The intersection over union (IoU) ratio between the candidate anchor box and the real box is calculated, and the candidate anchor boxes whose IoU exceeds the preset threshold are taken as positive samples, and the rest are taken as negative samples.

[0018] As a further solution of the present invention, for each scale in the multi-scale, a channel attention operation is performed on the initial feature map to obtain a channel attention map, and the channel attention weighted map is fused with the initial feature map. Specifically:

[0019] The target detection model uses a convolutional neural network as the basic framework to extract multi-scale features from the input satellite remote sensing image blocks. Through convolution kernels or pooling operations of different sizes, the feature representation of the image at different scales is captured to generate a multi-scale initial feature map.

[0020] Perform global average pooling on the initial feature map output by the convolutional layer, compressing the feature map of each channel into a real number with a global receptive field, representing the global information of the feature map of the channel, that is, a real number vector representing global information, whose dimension matches the number of channels of the input feature map;

[0021] The output of the global average pooling is reduced in dimension through a fully connected layer to reduce the amount of calculation and capture the nonlinear interaction between channels. The output of the fully connected layer is nonlinearly transformed using the ReLU activation function to introduce nonlinear factors.

[0022] Through another fully connected layer, the dimension of the initial feature map is restored to the same number of channels as the input feature map. The sigmoid activation function is used to transform the output after dimensionality increase, maintaining the weight value of each channel between 0 and 1, and outputting the channel attention map.

[0023] The obtained channel attention map is multiplied element-by-element with the initial feature map, and the channel attention is weighted to obtain the channel attention weighted map.

[0024] As a further solution of the present invention, the target detection result of each block image is fused to obtain the target detection result of the satellite remote sensing image, which specifically includes:

[0025] The dataset is divided into training set, validation set and test set. The target detection model is trained using the training set based on the improved YOLOv8 framework. The performance of the model is evaluated using the validation set, and the model is optimized using the cross entropy loss function and CIoU loss function.

[0026] Input the channel attention weighted map into the detection head for object classification and position regression;

[0027] Define a loss function to evaluate the difference between the model's predicted values ​​and the true values, and optimize the model parameters by minimizing the loss during training;

[0028] Iterative training is performed on the training set, and in each iteration, a forward propagation and a back propagation are performed on each sample in the training set to update the model parameters.

[0029] As a further solution of the present invention, the loss function is defined as follows:

[0030] The CIoU loss function formula is adjusted, and the ratio of the absolute difference to the diagonal length of the real box is used to deal with the diagonal length difference penalty term, specifically:

[0031]

[0032] IoU: represents the intersection over union of the predicted box and the true box;

[0033] ρ(b,bgt): represents the square of the Euclidean distance between the center point of the predicted box and the center point of the true box, which is used to measure the consistency of the center point.

[0034] c: is the diagonal length of the minimum enclosing box, which is the smallest rectangle that can contain both the predicted box and the real box, and is used to normalize the center point distance;

[0035] v: is the aspect ratio consistency measure, which together with IoU and center point distance determines the shape similarity of the predicted box;

[0036] α: is a balancing parameter used to adjust the weight of aspect ratio consistency in the loss function;

[0037] A, Agt: the areas of the predicted box and the true box respectively;

[0038] β: is the weight of the area ratio penalty term, which is used to control the contribution of area difference to the loss function;

[0039] γ: is the weight of the penalty term for the difference in diagonal length, which is used to control the contribution of shape difference to the loss function;

[0040] where ω and h are the width and height of the predicted bounding box, respectively, and gt and hgt are the width and height of the true bounding box, respectively.

[0041] As a further solution of the present invention, the target detection model is trained using data enhancement technology, a cross-validation method is used to select hyperparameters, and overfitting is prevented by setting an early stopping strategy. The data enhancement technology includes but is not limited to image rotation, flipping and scaling operations.

[0042] Another object of the present invention is to provide a remote sensing image aircraft target detection system integrating channel attention, the system comprising:

[0043] An image acquisition module is used to acquire satellite remote sensing images to be detected through public data sets;

[0044] The target detection module is used to divide the satellite remote sensing image into blocks, and input each block image into the target detection model to obtain the target detection result of each block image output by the target detection model;

[0045] The result fusion module is used to fuse the target detection results of each block image to obtain the target detection result of the satellite remote sensing image.

[0046] The beneficial effects of the present invention are:

[0047] Through multi-scale feature extraction and feature fusion, the present invention can fully utilize image information at different scales and improve the accuracy of target detection.

[0048] The large-scale satellite remote sensing image is divided into multiple small-scale block images for processing by adopting the block processing method, which reduces the computational complexity and improves the detection efficiency.

[0049] The introduction of the self-attention mechanism enables the model to pay more attention to the key areas in the image, improving the robustness of the model to complex backgrounds and occlusions. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1A flow chart of a remote sensing image aircraft target detection method integrating channel attention provided by an embodiment of the present invention;

[0051] Figure 2 It is a CSE graph;

[0052] Figure 3 It is the C2f, SPPF and CSE diagram;

[0053] Figure 4 This is the overall network structure diagram;

[0054] Figure 5 A structural block diagram of a remote sensing image aircraft target detection system integrating channel attention provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0056] It is understood that the terms "first", "second", etc. used in this application may be used herein to describe various elements, but unless otherwise specified, these elements are not limited by these terms. These terms are only used to distinguish a first element from another element. For example, without departing from the scope of this application, a first xx script may be referred to as a second xx script, and similarly, a second xx script may be referred to as a first xx script.

[0057] like Figure 1 As shown, a remote sensing image aircraft target detection method integrating channel attention, the method comprising:

[0058] S101: Obtain a satellite remote sensing image to be detected through a public data set;

[0059] S102: Divide the satellite remote sensing image into blocks, and input each block image into a target detection model to obtain a target detection result of each block image output by the target detection model;

[0060] The target detection model is used to perform multi-scale feature extraction on each block image to obtain a multi-scale initial feature map, perform feature fusion on the multi-scale feature map to obtain a multi-scale fused feature map, perform self-attention extraction on the multi-scale fused feature map to obtain a multi-scale enhanced feature map, and perform target detection based on the multi-scale enhanced feature map to obtain the target detection result of each block image;

[0061] S103: Fusing the target detection results of each block image to obtain the target detection result of the satellite remote sensing image.

[0062] (1) Data augmentation

[0063] Preprocessing strategies such as horizontal flipping and 90-degree rotation of the image are performed. 90-degree rotation includes clockwise and counterclockwise rotation. These transformations can increase the diversity of the image because they simulate the way objects are presented at different angles. The rotated images help the model learn features from different perspectives, thereby improving the generalization ability of the model.

[0064] (2) Positive and negative sample generation

[0065] The positive and negative samples of any sample block image of the sample satellite remote sensing image are obtained based on the following steps: multiple anchor frames are generated by a clustering algorithm; the center distance between each anchor frame and its corresponding real frame is calculated, and a preset number of anchor frames with the smallest center distance are selected as candidate anchor frames; the intersection over union (IoU) between each candidate anchor frame and its corresponding real frame is calculated, and the candidate anchor frames whose IoU exceeds a preset threshold are taken as positive samples, and the remaining candidate anchor frames are taken as negative samples.

[0066] In the training process of remote sensing satellite image target detection, the accurate generation of positive and negative samples is crucial to the training effect of the model. In order to effectively generate positive and negative samples from any sample block image of the sample satellite remote sensing image, the present invention adopts the following steps:

[0067] Step 1: Anchor box generation

[0068] First, clustering algorithms (such as K-means, K-means++, etc.) are used to cluster the true boxes (i.e., the labeled target bounding boxes) of the training set to generate multiple anchor boxes. These anchor boxes represent the possible shapes and sizes of the target and are the basis for subsequent matching with the true boxes.

[0069] Step 2: Candidate anchor box selection

[0070] For each true box, calculate the center distance between all anchor boxes and the true box. The center distance can be calculated based on the center point coordinates of the anchor box and the true box. Then, according to the size of the center distance, select a preset number of anchor boxes with the smallest center distance as candidate anchor boxes. The purpose of this step is to find a set of anchor boxes with the closest shape and size for each true box for subsequent matching operations.

[0071] Step 3: Intersection over Union (IoU) calculation and positive and negative sample determination

[0072] For each candidate anchor box, the intersection over union (IoU) between it and the corresponding real box is calculated. IoU is an indicator to measure the degree of overlap between two boxes. The larger the value, the higher the degree of overlap between the two boxes. Then, the candidate anchor boxes whose IoU exceeds the preset threshold are used as positive samples, which will be used to train the model to identify the target. The candidate anchor boxes whose IoU does not exceed the preset threshold are used as negative samples, which will be used to train the model to distinguish non-target areas.

[0073] In summary, through the above steps, the present invention can effectively generate positive and negative samples from any sample block image of the sample satellite remote sensing image, providing strong support for subsequent model training.

[0074] (3) Multi-scale feature extraction

[0075] The target detection model performs multi-scale feature extraction on each block image to obtain a multi-scale initial feature map. Then, for each scale in the multi-scale, a channel attention operation is performed on the initial feature map to obtain a channel attention map, which is then fused with the initial feature map to obtain a channel attention weighted map.

[0076] In remote sensing satellite image target detection, multi-scale feature extraction is a crucial step because it can capture target information of different scales in the image. In order to further improve the detection performance, the present invention combines channel attention to enhance and optimize the multi-scale features.

[0077] Step 1: Multi-scale feature extraction, such as Figure 3 As shown in the figure, there are C2f (C2f feature extraction module), SPPF (fast spatial pyramid pooling module) and CSE (Channel Spatial Excitation)

[0078] First, the target detection model uses a convolutional neural network as the basic framework to extract multi-scale features from the input satellite remote sensing image blocks. Through convolution kernels of different sizes or pooling operations, the model can capture the feature representation of the image at different scales and generate a multi-scale initial feature map.

[0079] Step 2: Channel attention operation, such as Figure 2 As shown, this is the CSE (Channel Spatial Excitation) diagram

[0080] 1. Perform global average pooling on the output feature map of the convolutional layer to compress the feature map of each channel into a real number with a global receptive field, which represents the global information of the feature map of the channel. That is, a real number vector (or feature map) representing global information, whose dimension matches the number of channels of the input feature map.

[0081] 2. Perform dimensionality reduction on the output of the previous operation through a fully connected layer (also called a dimensionality reduction layer) to reduce the amount of computation and capture nonlinear interactions between channels. The output dimension of this layer is usually much smaller than the number of channels of the input feature map. Use the ReLU activation function to perform a nonlinear transformation on the output of the fully connected layer to introduce nonlinear factors.

[0082] 3. The dimension of the feature map is restored to the same number of channels as the input feature map through another fully connected layer (also called a dimension-raising layer). The output after dimension-raising is transformed using the Sigmoid activation function to ensure that the weight value of each channel is between 0 and 1. In this way, each weight value can be interpreted as the importance of the corresponding feature channel.

[0083] Step 3: Channel Attention Weighted Graph Fusion

[0084] The obtained channel attention map is element-wise multiplied (or weighted summed) with the initial feature map to achieve channel attention weighting, thus obtaining the channel attention weighted map. This step enhances the model's attention to important channel features and suppresses redundant information.

[0085] Through the above-mentioned method of fusing multi-scale feature extraction and attention mechanism, the remote sensing satellite image target detection model provided by the present invention can more effectively extract and utilize key information in the image, thereby improving the accuracy and efficiency of target detection.

[0086] 6. The method according to claim 5, characterized in that fusing the target detection results of each block image to obtain the target detection result of the satellite remote sensing image specifically comprises:

[0087] (4) Model training and verification, such as Figure 4 As shown, this is the overall network structure diagram (Backbone, Neck, Head)

[0088] The preprocessed data set is divided into training set, validation set and test set. The target detection model is trained using the training set based on the improved YOLOv8 framework, and the performance of the model is evaluated using the validation set. During the training process, optimization algorithms such as the cross entropy loss function can be used to optimize the model parameters.

[0089] Backbone network is responsible for feature extraction, Neck is responsible for multi-scale feature fusion, and Head is responsible for the final target detection and classification tasks, including predicting bounding boxes, categories, and confidence levels.

[0090] C2f-feature extraction module, CSE-convolution + attention mechanism module, SPPF-fast spatial pyramid pooling module, Split-channel segmentation module, Bottleneck-bottleneck residual module, Concat-cascade module, MaxPool-maximum pooling module, Upsample-upsampling module.

[0091] The fused feature map is input into the detection head for target classification and position regression. The classification branch in the detection head is responsible for predicting the probability that each anchor belongs to each category, while the regression branch is responsible for adjusting the position and size of the anchor to make it closer to the actual target bounding box.

[0092] Define the loss function: Object detection usually includes classification loss (such as cross entropy loss) and positioning loss (complete intersection-over-union loss, CIoU Loss). The loss function is a function that evaluates the difference between the model's predicted value and the true value. The model parameters are optimized by minimizing the loss during training.

[0093] The CIoU loss function formula has been adjusted, especially in the treatment of the diagonal length difference penalty term, the ratio of the absolute difference to the diagonal length of the true box is used, because it can more directly reflect the similarity in shape between the predicted box and the true box, and does not depend on the absolute size of the box.

[0094]

[0095] IoU: represents the intersection over union of the predicted box and the true box;

[0096] ρ(b,bgt): represents the square of the Euclidean distance between the center point of the predicted box and the center point of the true box, which is used to measure the consistency of the center point.

[0097] c: is the diagonal length of the minimum enclosing box, which is the smallest rectangle that can contain both the predicted box and the real box, and is used to normalize the center point distance;

[0098] v: is the aspect ratio consistency measure, which together with IoU and center point distance determines the shape similarity of the predicted box;

[0099] α: is a balancing parameter used to adjust the weight of aspect ratio consistency in the loss function;

[0100] A, Agt: the areas of the predicted box and the true box respectively;

[0101] β: is the weight of the area ratio penalty term, which is used to control the contribution of area difference to the loss function;

[0102] γ: is the weight of the penalty term for the difference in diagonal length, which is used to control the contribution of shape difference to the loss function;

[0103] where ω and h are the width and height of the predicted bounding box, respectively, and gt and hgt are the width and height of the true bounding box, respectively.

[0104] In this embodiment, the target detection model is trained using data enhancement technology, a cross-validation method is used to select hyperparameters, and an early stopping strategy is set to prevent overfitting. The data enhancement technology includes but is not limited to image rotation, flipping and scaling operations.

[0105] The remote sensing satellite image target detection method and system provided by the present invention realizes rapid and accurate detection of targets in satellite remote sensing images by constructing a target detection model based on multi-scale feature extraction and self-attention mechanism. The method and system have the advantages of high detection accuracy, high detection efficiency, strong robustness, etc., and can be widely used in remote sensing image processing, environmental monitoring, urban planning and other fields.

[0106] like Figure 5 As shown, a remote sensing image aircraft target detection system integrating channel attention, the system comprises:

[0107] An image acquisition module is used to acquire satellite remote sensing images to be detected through public data sets;

[0108] The target detection module is used to divide the satellite remote sensing image into blocks, and input each block image into the target detection model to obtain the target detection result of each block image output by the target detection model;

[0109] The result fusion module is used to fuse the target detection results of each block image to obtain the target detection result of the satellite remote sensing image.

[0110] It should be understood that, although each step in the flow chart of each embodiment of the present invention is shown in sequence according to the indication of the arrow, these steps are not necessarily performed in sequence according to the order indicated by the arrow. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be performed in other orders. Moreover, at least a portion of the steps in each embodiment may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these sub-steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a portion of other steps or sub-steps or stages of other steps.

[0111] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0112] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

[0113] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A remote sensing image aircraft target detection method integrating channel attention, characterized in that: The method comprises: S101: Obtain a satellite remote sensing image to be detected through a public data set; S102: Divide the satellite remote sensing image into blocks, and input each block image into a target detection model to obtain a target detection result of each block image output by the target detection model; S103: Fusing the target detection results of each block image to obtain the target detection result of the satellite remote sensing image.

2. The method according to claim 1, characterized in that The target detection model is used to perform multi-scale feature extraction on each block image to obtain a multi-scale initial feature map, perform feature fusion on the multi-scale feature map to obtain a multi-scale fused feature map, perform self-attention extraction on the multi-scale fused feature map to obtain a multi-scale enhanced feature map, and perform target detection based on the multi-scale enhanced feature map to obtain the target detection result of each block image.

3. The method according to claim 1, characterized in that In S102, it also includes: Apply preprocessing strategy to the image; Obtain positive and negative samples of any sample block image of the sample satellite remote sensing image; For each scale in the multi-scale, a channel attention operation is performed on the initial feature map to obtain a channel attention map, which is then fused with the initial feature map to obtain a channel attention weighted map.

4. The method according to claim 3, characterized in that The positive and negative samples of any sample block image of the sample satellite remote sensing image are obtained, specifically: Use the clustering algorithm to cluster the real frames of the training set and generate multiple anchor frames; Calculate the center distance between each real box and all anchor boxes, and select a preset number of anchor boxes with the smallest center distance as candidate anchor boxes The intersection over union (IoU) ratio between the candidate anchor box and the real box is calculated, and the candidate anchor boxes whose IoU exceeds the preset threshold are taken as positive samples, and the rest are taken as negative samples.

5. The method according to claim 4, characterized in that For each scale in the multi-scale, a channel attention operation is performed on the initial feature map to obtain a channel attention map, and the channel attention weighted map is fused with the initial feature map. Specifically: The target detection model uses a convolutional neural network as the basic framework to extract multi-scale features from the input satellite remote sensing image blocks. Through convolution kernels or pooling operations of different sizes, the feature representation of the image at different scales is captured to generate a multi-scale initial feature map. Perform global average pooling on the initial feature map output by the convolutional layer, compressing the feature map of each channel into a real number with a global receptive field, representing the global information of the feature map of the channel, that is, a real number vector representing global information, whose dimension matches the number of channels of the input feature map; The output of the global average pooling is reduced in dimension through a fully connected layer to reduce the amount of calculation and capture the nonlinear interaction between channels. The output of the fully connected layer is nonlinearly transformed using the ReLU activation function to introduce nonlinear factors. Through another fully connected layer, the dimension of the initial feature map is restored to the same number of channels as the input feature map. The sigmoid activation function is used to transform the output after dimensionality increase, maintaining the weight value of each channel between 0 and 1, and outputting the channel attention map. The obtained channel attention map is multiplied element-by-element with the initial feature map, and the channel attention is weighted to obtain the channel attention weighted map.

6. The method according to claim 5, characterized in that The step of fusing the target detection results of each block image to obtain the target detection result of the satellite remote sensing image specifically includes: The dataset is divided into training set, validation set and test set. The target detection model is trained using the training set based on the improved YOLOv8 framework. The performance of the model is evaluated using the validation set, and the model is optimized using the cross entropy loss function and CIoU loss function. Input the channel attention weighted map into the detection head for object classification and position regression; Define a loss function to evaluate the difference between the model's predicted values ​​and the true values, and optimize the model parameters by minimizing the loss during training; Iterative training is performed on the training set, and in each iteration, a forward propagation and a back propagation are performed on each sample in the training set to update the model parameters.

7. The method according to claim 6, characterized in that The loss function is defined as follows: The CIoU loss function formula is adjusted, and the ratio of the absolute difference to the diagonal length of the real box is used to deal with the diagonal length difference penalty term, specifically: IoU: represents the intersection over union of the predicted box and the true box; ρ(b,bgt): represents the square of the Euclidean distance between the center point of the predicted box and the center point of the true box, which is used to measure the consistency of the center point. c: is the diagonal length of the minimum enclosing box, which is the smallest rectangle that can contain both the predicted box and the real box, and is used to normalize the center point distance; v: is the aspect ratio consistency measure, which together with IoU and center point distance determines the shape similarity of the predicted box; α: is a balancing parameter used to adjust the weight of aspect ratio consistency in the loss function; A, Agt: the areas of the predicted box and the true box respectively; β: is the weight of the area ratio penalty term, which is used to control the contribution of area difference to the loss function; γ: is the weight of the penalty term for the difference in diagonal length, which is used to control the contribution of shape difference to the loss function; where ω and h are the width and height of the predicted bounding box, respectively, and gt and hgt are the width and height of the true bounding box, respectively.

8. The method according to claim 7, characterized in that The target detection model is trained using data enhancement technology, a cross-validation method is used to select hyperparameters, and an early stopping strategy is set to prevent overfitting. The data enhancement technology includes but is not limited to image rotation, flipping and scaling operations.

Images