Method for rapidly and accurately detecting small ground dense targets by high-altitude high-speed aircraft

Abstract

The invention belongs to the field of cross fusion of the technical fields of remote sensing detection, aerospace and target detection, and relates to a method for a high-altitude high-speed aircraft to quickly and accurately detect dense small targets on the ground, which comprises the following steps of: firstly, refining an input feature map by using a multi-scale edge enhancement module, and extracting multi-scale features and enhancing edge information to obtain a multi-scale edge enhancement module; and the feature expression capability of the small ground target is improved. Thirdly, an efficient context attention mechanism is introduced, and through global information modeling, the calculation complexity is reduced, and meanwhile, the significance expression of small target features is optimized; and finally, performing dynamic up-sampling on feature maps of different scales by adopting a dynamic enhanced pyramid fusion technology to realize generation of a high-detail feature map, and enhancing the perception ability of the model to dense small targets at the same time.

Description

Technical Field

[0001] The present invention belongs to the cross-integration field of remote sensing detection, aerospace and target detection technology, and relates to a method for a high-altitude, high-speed aircraft to quickly and accurately detect densely packed small targets on the ground. Background Art

[0002] With the development of high-altitude, high-speed aircraft technology, camera-equipped high-altitude, high-speed aircraft (such as industrial and consumer drones) have been widely used in agricultural monitoring, aerial photography, logistics and transportation, security monitoring, and other fields. As a result, the demand for automatic understanding of visual data collected from these platforms is increasing, making the connection between computer vision and high-altitude, high-speed aircraft technology increasingly close.

[0003] Currently, deep learning-based algorithms have made significant progress in the field of object detection. Algorithms such as Faster R-CNN, SNIPER, YOLO, SSD, PANet, DETR, and RT-DETR can, to a certain extent, address the technical challenges of object detection. However, due to factors such as the relatively simple shooting angle of aerial images, the small size of targets in aerial images, extremely complex and diverse backgrounds, dense and uneven target distribution, diverse target scales, and target occlusion, traditional deep learning algorithms have low object detection accuracy in aerial images, making them difficult to truly apply in complex aerial imagery environments. Summary of the Invention

[0004] In order to solve the above technical problems existing in the prior art, the present invention proposes a method for quickly and accurately detecting dense small targets on the ground by a high-altitude, high-speed aircraft. The specific technical solution is as follows: A method for quickly and accurately detecting densely packed small targets on the ground by a high-altitude, high-speed aircraft, comprising: Step 1: Obtain a high-altitude, high-speed aircraft aerial photography dataset and align the annotation format of each aerial image; Step 2: Input the aerial image into the model, use the multi-scale edge enhancement module in the backbone network P2, P3, P4, and P5 layers of the model, and extract the multi-scale feature map through the multi-scale edge enhancement module; Step 3: Input the P5 layer feature map into the efficient context attention module, and generate an optimized feature matrix through normalization and global context vector calculation; Step 4: Dynamically upsample the optimized feature matrix and perform pyramid feature fusion with the dynamically upsampled P2, P3, and P4 feature maps; Step 5: Decode and output the fused image, and obtain the optimal model parameters through model training and evaluation, and finally output the final predicted box, category and detection probability of the target in the image.

[0005] Furthermore, the step 2 specifically includes: Step 2.1: Take the aerial image as input, pass it through the first convolutional layer of the backbone network, and output a Figure 1 / 2 size feature map; Step 2.2: After the feature map obtained in step 2.1 passes through the second layer of convolution, a feature map with a size of 1 / 4 of the original image is generated; Step 2.3: Pass the feature map of step 2.2 through the multi-scale edge enhancement module of the P2 layer to enhance the edge features; Step 2.4: After passing through the convolutional layer of the backbone network, continue to reduce the size of the feature map by 1 / 2; Step 2.5: Multi-scale edge enhancement is performed again by the multi-scale edge enhancement module of the P3 layer; Step 2.6: After passing through the convolutional layer of the backbone network, continue to reduce the size of the feature map by 1 / 2; Step 2.7: Multi-scale edge enhancement is performed again by the multi-scale edge enhancement module of the P4 layer; Step 2.8: After passing through the convolutional layer of the backbone network, continue to reduce the size of the feature map by 1 / 2; Step 2.9: Multi-scale edge enhancement is performed again after passing through the multi-scale edge enhancement module of the P5 layer.

[0006] Furthermore, the multi-scale edge enhancement module first uses an adaptive average pooling layer to scale the original input feature map into multiple feature maps of different scales; Then the number of feature map channels is reduced through a 1×1 convolution layer, and the enhanced feature map is generated through upsampling and edge enhancement operations; Finally, the original input and enhanced feature maps are concatenated and fused according to the channel dimension to generate the final edge-enhanced feature map.

[0007] Furthermore, in step 3, the efficient context attention module maps the input P5 layer feature map into a query matrix and a key matrix, and performs L2 norm normalization on the key matrix; The query matrix is weighted pooled using the global attention weights, and the columns of the weighted pooled query matrix are accumulated to generate a global context vector, which is then element-wise multiplied with the key matrix to generate a global context interaction matrix. The global context interaction matrix is linearly transformed, then added to the normalized query matrix, and the result of the addition is linearly transformed to finally output the optimized feature matrix.

[0008] Furthermore, the step 4 specifically includes: Step 4.1: Input the optimized feature matrix into the dynamic enhancement pyramid fusion module; Step 4.2: The dynamic enhancement pyramid fusion module dynamically upsamples the input features, performs 1×1 convolution, batch normalization, and SiLU processing; Step 4.3: Fuse the processing result of step 4.2 with the output of the multi-scale edge enhancement module of the P4 layer; Step 4.4: The fused result of step 4.3 is further processed using dynamic upsampling, 1×1 convolution, batch normalization, and SiLU. Step 4.5: Fuse the processing result of step 4.4 with the output of the multi-scale edge enhancement module of the P3 layer; Step 4.6: The fused result of step 4.5 is further processed using dynamic upsampling, 1×1 convolution, batch normalization, and SiLU. Step 4.7: Fuse the processing result of step 4.4 with the output of the multi-scale edge enhancement module of the P2 layer; Step 4.8: Downsample the fused result from step 4.7 using 3×3 convolution, batch normalization, and SiLU. Step 4.9: Fuse the downsampling result of step 4.8 with the fusion result of step 4.5 again; Step 4.10: Downsample the result of step 4.9 using 3×3 convolution, batch normalization, and SiLU; Step 4.11: Fuse the downsampling result of step 4.10 with the fusion result of step 4.3 again; Step 4.12: Downsample the result of step 4.11 using 3×3 convolution, batch normalization, and SiLU; Step 4.13: Fuse the downsampling result of step 4.12 with the optimized feature matrix; Step 4.14: Concatenate the results of steps 4.7, 4.9, 4.11, and 4.13 along the channel direction to obtain a comprehensive fusion feature map.

[0009] Furthermore, the dynamic upsampling process specifically includes: First, a 1×1 convolution is used to linearly transform the original feature map to generate a dynamic offset whose shape is [2 * S 2 , H, W], S represents the upsampling ratio, indicating the multiple that needs to be magnified, and H and W represent the height and width of the feature map; Then multiply the dynamic offset by 0.25, rearrange the sampling points of the dynamic offset, and combine the sampling points at the same position on the channel into a square S * S grid in sequence. The original [H, W] grid becomes a [S * H, S * W] grid. Regenerate the initial offset coordinates to represent the relative positions of the basic sampling grid points; Then, the rearranged dynamic offset is added to the initial offset coordinates to obtain the final sampling point coordinates; Finally, the original feature map is grid sampled according to the sampling point coordinates to generate an upsampled feature map with a size of [S * H, S * W].

[0010] Furthermore, the step 5 specifically includes: Step 5.1: After the integrated fusion feature map is selected through the intersection-over-union perceptual query, the decoder and the detection head, the model outputs the bounding box and category prediction; Step 5.2: Simultaneously conduct model training. After each round of model training, evaluate the effect using the validation set images and automatically save the model parameters with the best detection performance. Step 5.3: Using the trained model, output all possible final prediction boxes in the image as well as the categories and probabilities of the predicted targets.

[0011] The beneficial effects of the present invention are as follows: by improving the multi-scale edge enhancement module of the backbone network, the model's ability to recognize target edges is improved, thereby improving the feature extraction ability of small targets; adopting an efficient contextual attention mechanism, while reducing computational complexity, maintaining a high feature extraction effect; introducing dynamic upsampling technology, improving the quality of feature upsampling in pyramid feature fusion, and enhancing the model's ability to detect small targets; compared with the RT-DETR model, the average accuracy of the present invention in small target detection tasks is improved by 5.4%, and the number of parameters is reduced by 4.16M compared with the RT-DETR model. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 Schematic diagram of the overall structure of the DAE-DETR model for the method of quickly and accurately detecting densely packed small targets on the ground by a high-altitude, high-speed aircraft provided in an embodiment of the present invention; Figure 2 A schematic structural diagram of a multi-scale edge enhancement module provided in an embodiment of the present invention; Figure 3 A schematic diagram of the structure of an efficient contextual attention module provided by an embodiment of the present invention; Figure 4 A schematic diagram of the structure of a dynamic upsampling module provided in an embodiment of the present invention; Figure 5 and Figure 6 A schematic diagram comparing the small target detection effects of the DAE-DETR model and the RT-DETR model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0013] In order to make the purpose, technical solution and technical effect of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0014] This embodiment provides a method for quickly and accurately detecting densely packed small targets on the ground by a high-altitude, high-speed aircraft, including: Step 1: Obtain and preprocess the dataset: Obtain a high-altitude, high-speed aircraft aerial photography dataset and align the dataset's annotation format to suit training requirements. This includes: Step 1.1: Collect aerial image data taken by high-altitude, high-speed aircraft. This example uses the VisDrone2019 dataset, which was collected using high-altitude, high-speed aircraft of different models in different scenes, weather, and lighting conditions.

[0015] Step 1.2: Obtain or create the annotation files for these image data. To ensure the uniformity of training and the adaptability of the model, the annotation format of each image needs to be standardized. In the implementation process of this invention, the program automatically standardizes the annotation format of each image.

[0016] Step 1.3: Store the generated label file in the label folder and the collected image data in the image folder.

[0017] Step 1.4: Set the training path to the path of the label file and image file, set the batch size to 8, and the number of training rounds to 200 to start model training.

[0018] Step 2: Feature extraction through the backbone network with a multi-scale edge enhancement module: The backbone network uses a multi-scale edge enhancement module (MSEE) to enhance edge features at specific feature layers (P2, P3, P4, P5), such as Figure 2 As shown in the figure, a multi-scale edge enhancement module is used to extract feature maps and output pyramid feature maps P2, P3, P4 and P5. The recognition ability of small objects in the feature maps is improved through adaptive pooling and edge enhancement operations. Specifically, it includes: Step 2.1: Take the aerial image as input, pass it through the first convolution layer, and output a Figure 1 / 2 size feature map.

[0019] Step 2.2: After the second convolution layer, a feature map with a size of 1 / 4 of the original image is generated.

[0020] Step 2.3: The edge features are enhanced by MSEE in the P2 layer. The structure is shown in FIG2 .

[0021] The multi-scale edge enhancement module first uses four adaptive average pooling layers and a 1x1 convolution layer to generate feature maps of 3x3, 6x6, 9x9, 12x12, and the original size, respectively. The number of channels in the feature maps of the original size remains unchanged, while the number of channels in the feature maps of the remaining sizes is halved. The 3x3, 6x6, 9x9, and 12x12 feature maps are then sequentially subjected to 1x1 convolution, 3x3 convolution, and upsampling, ultimately returning them to the original size. These four feature maps are then processed by the edge enhancement layer. The edge enhancement layer performs average pooling on the input feature map, then subtracts the average pooled feature map from the input feature map. The result of this subtraction is then subjected to a 1x1 convolution to output an edge feature map, capturing edge information. The edge feature map is then added to the input feature map to generate the edge-enhanced feature map. After passing through the edge enhancement layer, these four feature maps are directly concatenated with the original feature map after 1x1 convolution, channel-wise, to form a feature map with the same size as the input image and twice the number of channels. Finally, the concatenated feature maps are fused using 1x1 convolution to produce the final edge-enhanced feature map, whose number of channels and size match the input image. The feature map enhanced by MSEE is not only passed down the backbone network but also serves as input to the dynamic enhancement pyramid fusion module.

[0022] Step 2.4: After passing through the convolutional layer of the backbone network, continue to reduce the size of the feature map by 1 / 2.

[0023] Step 2.5: Multi-scale edge enhancement is performed again after the MSEE of the P3 layer.

[0024] Step 2.6: After passing through the convolutional layer of the backbone network, continue to reduce the size of the feature map by 1 / 2.

[0025] Step 2.7: Multi-scale edge enhancement is performed again after the MSEE of the P4 layer.

[0026] Step 2.8: After passing through the convolutional layer of the backbone network, continue to reduce the size of the feature map by 1 / 2.

[0027] Step 2.9: Multi-scale edge enhancement is performed again after the MSEE of the P5 layer.

[0028] Through the above steps, the model can more fully extract the edge features of the target, especially the detailed features of small targets, thereby improving the accuracy and robustness of target detection.

[0029] Step 3: If Figure 3 As shown in the figure, the Efficient Contextual Attention (ECA) module is used to further extract information from the P5 feature map: the ECA module uses the ECA mechanism to generate an optimized feature matrix through normalization and global context vector calculation, reducing the computational complexity of the traditional self-attention mechanism while retaining the global context information. Specifically, it includes: Step 3.1: Input the initial feature matrix Using the projection weight matrix Perform linear transformation to obtain the query matrix Q and bond matrix K : , .

[0030] Step 3.2: The matrix is normalized by L2 norm to ensure that the relative relationship between different vectors is more stable.

[0031] Step 3.3: Using a Learnable Vector , the global attention weight vector is calculated by the following formula : , in, , represents the global attention weight of each input feature.

[0032] Step 3.4: According to Attention weights to query matrix Perform weighted pooling, i.e. Each column of the matrix Multiply by the corresponding weight . The query matrix after weighted pooling The columns are accumulated to obtain the global context vector : , in, represent No. List, Represents the query matrix No. List, Represents a matrix multiplication operation.

[0033] Step 3.5: Transform the global context vector Broadcast and expand its shape to The matrix and the normalized key matrix Multiply element by element to get the global context interaction matrix .

[0034] Step 3.6: Perform a linear transformation of the 1x1 convolution and then add the normalized query matrix , and then perform linear transformation on the added result using 1x1 convolution, and finally output the updated feature matrix.

[0035] Traditional self-attention can effectively model the interaction between input tokens. Specifically, the input of traditional self-attention is ,in , contains n tokens, each token is a d-dimensional embedding vector. Input Through three matrices 、 、 is projected as the query matrix Q , key matrix K Sum Matrix V Each self-attention layer consists of h This allows the model to focus on different aspects of the input. The self-attention mechanism can be described as: , in n is the number of tokens, d is the hidden dimension. In general, the complexity of the self-attention mechanism is .

[0036] For longer sequences, the inference speed of the self-attention mechanism will be slower and the memory usage will be higher, so it is not practical in scenarios with high real-time requirements.

[0037] This mechanism optimizes the complex calculation between the query and key matrix and discards the value matrix V The related calculations of attention are carried out to achieve the purpose of improving the efficiency and results of attention calculation.

[0038] Step 4: Reference Figure 1 and Figure 4, dynamically enhance the pyramid fusion of feature maps: generate high-quality feature maps through the dynamic upsampling module, and perform pyramid feature fusion with other dynamically upsampled feature maps to improve the model's performance in small target detection and edge feature extraction. Then, the fused multi-layer feature maps are spliced to generate the final comprehensive fused feature map. Specifically including: Step 4.1: Use the optimized feature matrix as the input feature map of the dynamic enhancement pyramid fusion module DFPF.

[0039] Step 4.2: The dynamic enhancement pyramid fusion module DFPF will first dynamically upsample the feature map. The structure of dynamic upsampling is shown in the attached figure. Figure 4 As shown. The dynamic upsampling module first uses a 1×1 convolution to linearly transform the original feature map to generate a dynamic offset whose shape is [2 * S 2 , H, W]. S represents the upsampling ratio, which indicates the magnification factor required. The upsampling ratio in this embodiment is 2. H and W represent the height and width of the feature map. The dynamic offset is then multiplied by 0.25 to reduce the offset amplitude of the sampling point and ensure the stability of dynamic sampling. The sampling points of the dynamic offset are rearranged, and the sampling points at the same position on the channel are sequentially combined into a square S * S grid, so that the original [H, W] grids become [S * H, S * W] grids. The initial offset coordinates are then generated to indicate the relative position of the basic sampling grid points. The rearranged dynamic offset is added to the initial offset coordinates to obtain the final sampling point coordinates. The original feature map is grid sampled according to the sampling point coordinates to generate an upsampled feature map with a size of [S * H, S * W]. It is then further processed through 1×1 convolution, batch normalization, and SiLU.

[0040] Step 4.3: Fuse the processing result of step 4.2 with the output of the P4 layer MSEE.

[0041] Step 4.4: The fused results from step 4.3 are further processed using dynamic upsampling, 1×1 convolution, batch normalization, and SiLU.

[0042] Step 4.5: Fuse the processing result of step 4.4 with the output of the P3 layer MSEE.

[0043] Step 4.6: The fused results from step 4.5 are further processed using dynamic upsampling, 1×1 convolution, batch normalization, and SiLU.

[0044] Step 4.7: Fuse the processing result of step 4.4 with the output of the P2 layer MSEE.

[0045] Step 4.8: Downsample the result of step 4.7 using 3×3 convolution, batch normalization, and SiLU.

[0046] Step 4.9: The downsampling result of step 4.8 is once again fused with the fusion result of step 4.5.

[0047] Step 4.10: Downsample the result of step 4.9 using 3×3 convolution, batch normalization, and SiLU.

[0048] Step 4.11: The downsampling result of step 4.10 is once again fused with the fusion result of step 4.3.

[0049] Step 4.12: Downsample the result of step 4.11 using 3×3 convolution, batch normalization, and SiLU.

[0050] Step 4.13: Perform feature fusion on the downsampling result of step 4.12 and the result after ECA processing of step 4.1.

[0051] Step 4.14: Splice the results of steps 4.7, 4.9, 4.11, and 4.13 along the channel direction.

[0052] This embodiment abandons the traditional convolution kernel-based upsampling and adopts dynamic upsampling technology. This technology achieves high-quality upsampling of feature maps through an adaptive point segmentation mechanism, improving the ability to retain edge details. The high-resolution feature map generated by dynamic upsampling is fully integrated with the features of each layer of the pyramid, achieving clearer edge capture of the feature map during the upsampling process, improving the detection accuracy of small targets, and enhancing the edge feature extraction capability of the model. Dynamic upsampling amplifies the semantic information of the deep feature map with high quality, and pyramid feature fusion merges the shallow detail information with the deep semantic information after high-quality amplification. Through the interactive fusion of shallow and deep feature information, the performance of the model in small target detection tasks is enhanced, and the detection accuracy and stability of small targets are greatly improved.

[0053] Step 5: Train the model and output the detection results: Load the optimal model parameters and output the final predicted box, category, and detection probability of the target in the image. Specifically, it includes: Step 5.1: After the concatenated feature map after feature fusion passes through the final IoU-aware query selection, decoder and detection head, the model will output the bounding box and category prediction.

[0054] Step 5.2: After each round of model training, the validation set images are evaluated for performance. Based on comprehensive evaluations of mAp50, mAp50-95, recall, and precision, the model parameters with the best detection performance are automatically saved. This selective saving feature ensures optimal parameters are obtained even when too many training rounds cause a decrease in model accuracy. Training is not performed every time the model is used; it is a prerequisite for the model to acquire detection capabilities. Once the model is trained, it can be used directly on images for all subsequent tests.

[0055] Step 5.3: Using the trained model, output all possible final prediction boxes in the image as well as the categories and probabilities of the predicted targets.

[0056] Combined with attachment Figure 5 、 Figure 6 The results of the improved model are illustrated with relevant data. To verify the model's performance, we used a batch size of 8 and 200 training rounds, and tested the model's performance through multiple comparative experiments and ablation experiments. The experimental results show that the DAE-DETR model improves mAP50 by 5.4% compared to the RT-DETR model, while reducing the number of parameters by 4.16M. The specific experimental data is shown in Table 1: Table 1 In order to demonstrate the effect achieved by the invention, Figure 5 and Figure 6 Description: Compare the detection images of the DAE-DETR model and the RT-DETR model. The left side is the detection image of the DAE-DETR model, and the right side is the detection image of the RT-DETR model. The image above the two images is the original image being detected. Figure 5 As can be seen from the right picture, under small target detection, the RT-DETR model has insufficient small target recognition ability and regards many small waste items lying on the ground as tricycles. Figure 5 In the image on the left, the DAE-DETR model is able to successfully reject these seemingly good-looking rejects. Figure 6 In the picture on the right, the RT-DETR model treats square objects that are about the same width as a person as pedestrians, while Figure 6 In the left picture, the DAE-DETR model can correctly exclude the square object, and in the given detection and classification confidence, the DAE-DETR model is better than the RT-DETR model; according to the comparison Figure 5 and 6 ,It can be seen that the detection performance of the DAE-DETR model in dense small targets is much better than that of the RT-DETR model.

[0057] In summary, this paper proposes a Dynamic Attention and Edge-enhanced Detection Transformer (DAE-DETR) algorithm that combines multi-scale feature fusion with edge enhancement technology to achieve superior small object detection performance compared to traditional models. This improved approach can better address challenges such as dense small objects, diverse object scales, and object occlusion, significantly improving object detection accuracy and robustness in complex aerial photography scenarios.

[0058] The above description is only a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the implementation process of the present invention is described in detail above, it is still possible for those familiar with the art to modify the technical solutions described in the above examples or to replace some of the technical features therein with equivalents. Any modifications, equivalent replacements, etc. made within the spirit and principles of the present invention shall be included in the scope of protection of the present invention.

Claims

1. A method for quickly and accurately detecting densely packed small targets on the ground by a high-altitude, high-speed aircraft, characterized in that: include: Step 1: Obtain a high-altitude, high-speed aircraft aerial photography dataset and align the annotation format of each aerial image; Step 2: Input the aerial image into the model, use the multi-scale edge enhancement module in the backbone network P2, P3, P4, and P5 layers of the model, and extract the multi-scale feature map through the multi-scale edge enhancement module; Step 3: Input the P5 layer feature map into the efficient context attention module, and generate an optimized feature matrix through normalization and global context vector calculation; Step 4: Dynamically upsample the optimized feature matrix and perform pyramid feature fusion with the dynamically upsampled P2, P3, and P4 feature maps; Step 5: Decode and output the fused image, and obtain the optimal model parameters through model training and evaluation, and finally output the final predicted box, category and detection probability of the target in the image.

2. The method for quickly and accurately detecting small targets densely packed on the ground by a high-altitude, high-speed aircraft as claimed in claim 1, characterized in that: The step 2 specifically includes: Step 2.1: Take the aerial image as input, pass it through the first convolutional layer of the backbone network, and output a feature map with a size of 1 / 2 the original image; Step 2.2: After the feature map obtained in step 2.1 passes through the second layer of convolution, a feature map with a size of 1 / 4 of the original image is generated; Step 2.3: Pass the feature map of step 2.2 through the multi-scale edge enhancement module of the P2 layer to enhance the edge features; Step 2.4: After passing through the convolutional layer of the backbone network, continue to reduce the size of the feature map by 1 / 2; Step 2.5: Multi-scale edge enhancement is performed again by the multi-scale edge enhancement module of the P3 layer; Step 2.6: After passing through the convolutional layer of the backbone network, continue to reduce the size of the feature map by 1 / 2; Step 2.7: Multi-scale edge enhancement is performed again by the multi-scale edge enhancement module of the P4 layer; Step 2.8: After passing through the convolutional layer of the backbone network, continue to reduce the size of the feature map by 1 / 2; Step 2.9: Multi-scale edge enhancement is performed again after passing through the multi-scale edge enhancement module of the P5 layer.

3. The method for quickly and accurately detecting small, densely packed targets on the ground by a high-altitude, high-speed aircraft as claimed in claim 2, characterized in that: The multi-scale edge enhancement module first uses an adaptive average pooling layer to scale the original input feature map into multiple feature maps of different scales; Then the number of feature map channels is reduced through a 1×1 convolution layer, and the enhanced feature map is generated through upsampling and edge enhancement operations; Finally, the original input and enhanced feature maps are concatenated and fused according to the channel dimension to generate the final edge-enhanced feature map.

4. The method for quickly and accurately detecting small, densely packed targets on the ground by a high-altitude, high-speed aircraft as claimed in claim 2, characterized in that: In step 3, the efficient context attention module maps the input P5 layer feature map into a query matrix and a key matrix, and performs L2 norm normalization on the key matrix; The query matrix is weighted and pooled using the global attention weights. The columns of the weighted pooled query matrix are then accumulated to generate a global context vector, which is then element-wise multiplied with the key matrix to generate a global context interaction matrix. The global context interaction matrix is linearly transformed, then added to the normalized query matrix, and the result of the addition is linearly transformed to finally output the optimized feature matrix.

5. The method for quickly and accurately detecting small, densely packed targets on the ground by a high-altitude, high-speed aircraft as claimed in claim 4, characterized in that: The step 4 specifically includes: Step 4.1: Input the optimized feature matrix into the dynamic enhancement pyramid fusion module; Step 4.2: The dynamic enhancement pyramid fusion module dynamically upsamples the input features, performs 1×1 convolution, batch normalization, and SiLU processing; Step 4.3: Fuse the processing result of step 4.2 with the output of the multi-scale edge enhancement module of the P4 layer; Step 4.4: The fused result of step 4.3 is further processed using dynamic upsampling, 1×1 convolution, batch normalization, and SiLU. Step 4.5: Fuse the processing result of step 4.4 with the output of the multi-scale edge enhancement module of the P3 layer; Step 4.6: The fused result of step 4.5 is further processed using dynamic upsampling, 1×1 convolution, batch normalization, and SiLU. Step 4.7: Fuse the processing result of step 4.4 with the output of the multi-scale edge enhancement module of the P2 layer; Step 4.8: Downsample the fused result from step 4.7 using 3×3 convolution, batch normalization, and SiLU. Step 4.9: Fuse the downsampling result of step 4.8 with the fusion result of step 4.5 again; Step 4.10: Downsample the result of step 4.9 using 3×3 convolution, batch normalization, and SiLU; Step 4.11: Fuse the downsampling result of step 4.10 with the fusion result of step 4.3 again; Step 4.12: Downsample the result of step 4.11 using 3×3 convolution, batch normalization, and SiLU; Step 4.13: Fuse the downsampling result of step 4.12 with the optimized feature matrix; Step 4.14: Concatenate the results of steps 4.7, 4.9, 4.11, and 4.13 along the channel direction to obtain a comprehensive fusion feature map.

6. The method for quickly and accurately detecting small, densely packed targets on the ground by a high-altitude, high-speed aircraft as claimed in claim 5, characterized in that: The dynamic upsampling process specifically includes: First, a 1×1 convolution is used to linearly transform the original feature map to generate a dynamic offset whose shape is [2 * S 2 , H, W], S represents the upsampling ratio, indicating the multiple that needs to be magnified, and H and W represent the height and width of the feature map; Then multiply the dynamic offset by 0.25, rearrange the sampling points of the dynamic offset, and combine the sampling points at the same position on the channel into a square S * S grid in sequence. The original [H, W] grid becomes a [S * H, S * W] grid. Regenerate the initial offset coordinates to represent the relative positions of the basic sampling grid points; Then, the rearranged dynamic offset is added to the initial offset coordinates to obtain the final sampling point coordinates; Finally, the original feature map is grid sampled according to the sampling point coordinates to generate an upsampled feature map with a size of [S *H, S * W].

7. The method for quickly and accurately detecting small, densely packed targets on the ground by a high-altitude, high-speed aircraft as claimed in claim 5, characterized in that: The step 5 specifically includes: Step 5.1: After the integrated fusion feature map is passed through the intersection-over-union (IoU) perceptual query selection, decoder, and detection head, the model outputs the bounding box and category prediction; Step 5.2: Simultaneously conduct model training. After each round of model training, evaluate the effect using the validation set images and automatically save the model parameters with the best detection performance. Step 5.3: Using the trained model, output all possible final prediction boxes in the image as well as the categories and probabilities of the predicted targets.

Images