An aerial remote sensing image target detection method based on an HSNet model

Abstract

The application discloses the technical field of deep learning, computer vision and remote sensing image processing, and particularly relates to an aerial remote sensing image target detection method based on an HSNet model, which comprises the following specific steps: S1: constructing an aerial remote sensing image target detection data set; S2: constructing a multi-scale context feature enhancement extractor; S3: constructing an adaptive feature refining and fusing device; S4: based on the multi-scale context feature enhancement extractor and the adaptive feature refining and fusing device, constructing an HSNet model for aerial remote sensing image target detection; and S5: based on the aerial remote sensing image target detection data set, training the HSNet model to obtain a training-optimized HSNet model. In the process of aerial remote sensing image target detection, the application effectively improves the small target detection capability, and simultaneously enhances the spatial position perception and multi-scale feature fusion performance.

Description

Technical Field

[0001] This invention relates to the fields of deep learning, computer vision, and remote sensing image processing technology, specifically to a target detection method for aerial remote sensing images based on the HSNet model. Background Technology

[0002] Target detection in aerial remote sensing images has significant application value in many key fields such as national defense and security, environmental monitoring, disaster assessment, and urban planning, providing crucial information support for real-time situational awareness and decision support. In recent years, deep learning-based target detection methods have achieved excellent performance in natural scene images; however, their detection performance remains significantly insufficient when dealing with aerial remote sensing images.

[0003] Aerial remote sensing images are typically taken at high altitudes, resulting in small-scale and densely distributed targets. They also suffer from low overall resolution, insufficient brightness, and poor contrast, leading to a loss of effective features for small targets and significantly increasing the difficulty of detection. Furthermore, external factors such as platform movement and atmospheric disturbances can cause target aliasing between the target and the background, further complicating accurate detection.

[0004] The core challenges of target detection in aerial remote sensing lie in the small and densely distributed target scale, weak feature representation capabilities, severe background interference, and the complexity arising from the coexistence of targets at multiple scales. Therefore, a high-precision, high-efficiency target detection method is urgently needed to address the unique challenges of aerial remote sensing images.

[0005] Currently, deep learning-based target detection methods for aerial remote sensing images can be divided into two categories: two-stage detection and single-stage detection. Compared with two-stage detection algorithms, single-stage detection algorithms have faster computation speed and lower accuracy loss, making them more advantageous in real-time detection tasks such as UAV inspections. The YOLO series of algorithms, as typical single-stage target detection algorithms, have been widely used in the field of remote sensing in recent years due to their speed and efficiency, and have become the basic framework for current research.

[0006] To address the unique characteristics of aerial remote sensing images, the YOLO series of target detection algorithms have been continuously improved and expanded in the field of aerial remote sensing image processing. In terms of feature enhancement, FFCA-YOLO incorporates a feature aggregation module to enhance multi-scale information fusion, while SFC-YOLOv5 combines semantic feature compensation with global receptive field fusion, thereby improving the discrimination ability and robustness of small targets. Regarding multi-scale processing, LSKF-YOLO proposes a large spatial convolutional selective attention and multi-scale alignment mechanism, while LSKAFF-YOLO employs a highly separable convolutional attention module and a progressive feature pyramid, effectively optimizing the detection performance of targets of different sizes. In terms of lightweight and real-time optimization, YOLO-Fine reduces the number of parameters by simplifying the backbone network, LMSFA-YOLO integrates a lightweight multi-scale convolutional and hybrid attention structure, FBRT-YOLO introduces a lightweight module to improve inference speed while maintaining accuracy, and RN-YOLO, combined with the RepGhostC2f structure, further reduces computational complexity.

[0007] Although existing target detection algorithms have made some progress in remote sensing image target detection, the following key problems still exist: First, small targets, due to their small pixel ratio and the limited receptive field of traditional convolution, have insufficient feature representation capabilities, making them prone to confusion and missed detection in dense or low-contrast scenes; second, conventional parameter-sharing convolution mechanisms are difficult to perceive spatial structural differences and lack the ability to model local context and key locations, resulting in inaccurate localization of irregular targets; third, multi-scale fusion strategies lack flexibility and cross-level interaction mechanisms, shallow details are easily lost, and small target features are easily suppressed by large targets during the fusion process, leading to scale imbalance and limiting overall detection performance. Therefore, this paper proposes a target detection method for aerial remote sensing images based on the HSNet model. Summary of the Invention

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0009] A target detection method for aerial remote sensing images based on the HSNet model includes the following specific steps:

[0010] S1: Construct a target detection dataset for aerial remote sensing images;

[0011] S2: Construct a multi-scale contextual feature enhancement extractor;

[0012] S3: Construct an adaptive feature refinement fusion engine;

[0013] S4: Based on a multi-scale contextual feature enhancement extractor and an adaptive feature refinement fusioner, an HSNet model is constructed for target detection in aerial remote sensing images;

[0014] S5: Based on the target detection dataset of aerial remote sensing images, the HSNet model is trained to obtain the trained and optimized HSNet model;

[0015] S6: The trained and optimized HSNet model is applied to the target detection scenario of aerial remote sensing images to achieve intelligent identification and localization of targets. Finally, the detection result image is obtained, which includes the target's corresponding category, the coordinates of the target bounding box, and the confidence score, providing reliable technical support and decision-making basis for remote sensing image analysis and other fields.

[0016] As a preferred embodiment of the aerial remote sensing image target detection method based on the HSNet model described in this invention, the specific steps of S1 are as follows:

[0017] S11: Based on existing public datasets, collect aerial remote sensing images and construct an initial target detection dataset;

[0018] S12: Preprocess the initial target detection dataset to obtain the preprocessed target detection dataset;

[0019] S13: Divide the preprocessed target detection dataset to obtain the aerial remote sensing image target detection dataset.

[0020] As a preferred embodiment of the aerial remote sensing image target detection method based on the HSNet model described in this invention, the specific construction process of the multi-scale contextual feature enhancement extractor in S2 is as follows:

[0021] S21: Input any aerial remote sensing target detection image F1 into the DHFaster module, and output the aerial remote sensing target detection feature map F2 by efficiently capturing refined spatial features and contextual information;

[0022] S22: Then F2 is input into the DHFaster module to further capture refined spatial features and contextual information, and outputs the airborne remote sensing target detection feature map F3;

[0023] S23: Then input F3 into the MCM module to enhance the feature extraction capability of the target and output the aerial remote sensing target detection feature map F4;

[0024] S24: Input F4 into the DHFaster module and output the aerial remote sensing target detection feature map F5. Then input F5 into the MCM module and output the aerial remote sensing enhancement feature map P1.

[0025] S25: Input P1 into the DHFaster module and output the aerial remote sensing target detection feature map F6. Then input F6 into the MCM module and output the aerial remote sensing enhancement feature map P2.

[0026] S26: Input P2 into the DHFaster module and output the aerial remote sensing target detection feature map F7. Then input F7 into the MCM module and output the aerial remote sensing target detection feature map F8.

[0027] S27: Finally, input F8 into the SPPF module to generate the aerial remote sensing enhanced feature map P3;

[0028] S28: After processing by the multi-scale contextual feature enhancement extractor described above, aerial remote sensing enhanced feature maps P1, P2 and P3 are obtained.

[0029] As a preferred embodiment of the aerial remote sensing image target detection method based on the HSNet model described in this invention, the construction and execution process of the DHFaster module in S21 is as follows:

[0030] S211: Treat the input feature map of the DHFaster module as a context-interactive fusion feature map H1, and process H1 through two branches:

[0031] In the first branch, H1 is first subjected to a Conv1×1 convolution operation to initially adjust the number of channels and retain the original spatial location information, resulting in a context-interactive fusion feature map H2; then, H2 is subjected to a Conv3×3 convolution operation to extract information on local image features, resulting in a context-interactive fusion feature map H3; finally, H3 is subjected to a Conv1×1 convolution operation to fuse features from different channels at the same location, resulting in a context-interactive fusion feature map H4.

[0032] In the second branch, H1 is first divided into three groups—Group1, Group2, and Group3—along the channel dimension, allowing different groups to focus on different types of feature patterns. Then, the feature map of each group is further divided into k×k non-overlapping local sub-regions. Each local sub-region uses an independent, non-shared convolutional kernel for feature extraction, ensuring the model can capture and preserve subtle feature differences at each unique location. Afterward, all processed, size-invariant sub-regions are reassembled in their original positions, restoring the feature map to its original size before sub-region division. After these operations, Group1... Group 1 yields context interaction fusion feature map H5_1, Group 2 yields context interaction fusion feature map H5_2, and Group 3 yields context interaction fusion feature map H5_3. Then, H5_1, H5_2, and H5_3 are concatenated along the channel dimension to generate context interaction fusion feature map H5. Next, H5 is convolved with a Conv1×1 convolution to achieve information exchange and fusion among the three groups, resulting in context interaction fusion feature map H6. Finally, H6 is convolved with a Conv3×3 convolution to generate context interaction fusion feature map H7.

[0033] S212: Finally, perform a Concat operation on H4 and H7 to complete the splicing along the channel dimension and generate a contextual interactive fusion feature map H8; H8 is the output feature map of the DHFaster module.

[0034] As a preferred embodiment of the aerial remote sensing image target detection method based on the HSNet model described in this invention, the construction and execution process of the MCM module in S23 is as follows:

[0035] S231: Treat the input feature map of the MCM module as a multi-domain enhanced fusion feature map X1, and process X1 through three branches:

[0036] In the first branch, X1 is first subjected to a Conv1×1 convolution operation to obtain a multi-domain enhanced fusion feature map X2; then X2 is subjected to a Conv3×3 convolution operation to extract local spatial features to obtain a multi-domain enhanced fusion feature map X3; then X3 is input into a dilated convolution module DilatedConv with a kernel size of 3×3 and a dilation rate of 1 to further enhance and consolidate the local feature expression to obtain a multi-domain enhanced fusion feature map X4; X1 and X4 are subjected to a Concat operation to complete the concatenation along the channel dimension to enrich the hierarchy of features and generate a multi-domain enhanced fusion feature map X5;

[0037] In the second branch, X1 is first subjected to a Conv1×1 convolution operation to obtain a multi-domain enhanced fusion feature map X6; then X6 is subjected to a Conv3×3 convolution operation to extract local spatial features, resulting in a multi-domain enhanced fusion feature map X7; X7 is then input into a dilated convolution module DilatedConv with a kernel size of 3×3 and a dilation rate of 3 to capture mid-range contextual information, resulting in a multi-domain enhanced fusion feature map X8; X1 and X8 are then subjected to a Concat operation to complete the concatenation along the channel dimension to enrich the hierarchical nature of the features, generating a multi-domain enhanced fusion feature map X9;

[0038] In the third branch, X1 is first subjected to a Conv1×1 convolution operation to obtain a multi-domain enhanced fusion feature map X10; then, X10 is subjected to a Conv3×3 convolution operation to extract local spatial features, resulting in a multi-domain enhanced fusion feature map X11; then, X11 is input into a DilatedConv convolution module with a kernel size of 3×3 and a dilation rate of 5 to capture long-range dependencies and global scene context, resulting in a multi-domain enhanced fusion feature map X12; finally, X1 and X12 are subjected to a Concat operation to complete the concatenation along the channel dimension to enrich the hierarchy of features, generating a multi-domain enhanced fusion feature map X13;

[0039] S232: Subsequently, through three learnable fusion weights Adaptive fusion is performed on X5, X9 and X13 to generate multi-domain enhanced fusion feature map X14;

[0040] S233: Then, perform a Conv1×1 convolution operation on X14 to adjust the number of channels of X14 to match the number of channels of X1, and obtain the multi-domain enhanced fusion feature map X15; finally, add X1 and X15 element by element to prevent the gradient from vanishing in the deep network, and obtain the multi-domain enhanced fusion feature map X16; X16 is the output feature map of the MCM module.

[0041] As a preferred embodiment of the aerial remote sensing image target detection method based on the HSNet model described in this invention, the specific construction process of the adaptive feature refinement fusion unit in S3 is as follows:

[0042] S31: First, input the enhanced feature map P3 of the airborne remote sensing obtained by the multi-scale context feature enhancement extractor into the DCF_L module to enhance the representation ability of large targets and overall structure, and output the airborne remote sensing target detection feature map F9.

[0043] S32: Then, upsample F9 to generate the aerial remote sensing target detection feature map F10; then, add the aerial remote sensing enhanced feature map P2 obtained by the multi-scale context feature enhancement extractor and F10 element by element to obtain the aerial remote sensing target detection feature map F11; input F11 into the C2f module (n=3) to obtain the aerial remote sensing target detection feature map F12; input F12 into the DCF_S module to further enhance the representation ability of small targets in the features, and output the aerial remote sensing target detection feature map F13; then, upsample F13 to generate the aerial remote sensing target detection feature map F14;

[0044] S33: Next, the aerial remote sensing enhanced feature maps P1 and F14 obtained by the multi-scale context feature enhancement extractor are added element by element to obtain the aerial remote sensing target detection feature map F15; F15 is input into the C2f module (n=3) to refine the target features and obtain the aerial remote sensing refined fusion feature map R1.

[0045] S34: Input R1 into the DHFaster module to further capture refined spatial features and contextual information, and output the aerial remote sensing target detection feature map F16; then add F13 and F16 element by element to obtain the aerial remote sensing target detection feature map F17.

[0046] S35: Input F17 into the C2f module (n=3) to further refine the target features and obtain the refined fusion feature map R2 of airborne remote sensing; input R2 into the DHFaster module to deeply capture refined spatial features and contextual information and output the airborne remote sensing target detection feature map F18.

[0047] S36: Then add F9 and F18 element by element to obtain the aerial remote sensing target detection feature map F19; then input F19 into the C2f module (n=3) to perform deep refinement of the target features and obtain the aerial remote sensing refined fusion feature map R3.

[0048] S37: After processing by the above adaptive feature refinement fusion processor, the refined and fused feature maps R1, R2 and R3 of airborne remote sensing are obtained.

[0049] As a preferred embodiment of the aerial remote sensing image target detection method based on the HSNet model described in this invention, the construction and execution process of the DCF_L module in S31 is as follows:

[0050] S311: Treat the input feature map of the DCF_L module as a large target cascaded fusion feature map Q1, input Q1 into the multi-branch redundant feature generation module for the large target feature map, and output the large target cascaded fusion feature map Q2.

[0051] S312: Q2 is then processed through two branches to further process and fuse the target features:

[0052] In the first branch, MaxPool is first performed on Q2, using a 3×3 pooling window that slides with a stride of 1 pixel and fills the edges of Q2 with a full ring of pixels to highlight salient features, resulting in the large target cascaded fusion feature map Q3. Simultaneously, AvgPool is performed on Q2, using a 3×3 pooling window that slides with a stride of 1 pixel and fills the edges of Q2 with a full ring of pixels to suppress noise and smooth features, resulting in the large target cascaded fusion feature map Q4. Then, Q3 and Q4 are concatenated along the channel dimension to obtain the large target cascaded fusion feature map Q5. Finally, a Conv1×1 convolution operation is performed on Q5 to make the features more compact, resulting in the large target cascaded fusion feature map Q6. Q6 is then input into Channel Attention. The Channel Attention Module learns to obtain importance weight vectors W1 for different channels, which are used to evaluate the importance of each feature channel. Then, based on W1, a weighting operation is performed on the channel dimensions of Q6, that is, W1 is multiplied by Q6 channel by channel to obtain the large target cascaded fusion feature map Q7.

[0053] In the second branch, Q2 is first subjected to a Conv3×3 convolution operation to further extract and transform the features, resulting in a large target cascaded fusion feature map Q8. Then, Q8 is subjected to a Conv1×1 convolution operation to obtain a large target cascaded fusion feature map Q9. Next, Q9 is input into the Spatial Attention Module to learn the importance weight matrix W2 of the feature map in the spatial dimension, which is used to evaluate the importance of each spatial position. Then, the spatial positions of Q9 are weighted according to W2, that is, W2 is multiplied with Q9 position by position to obtain a large target cascaded fusion feature map Q10.

[0054] S313: Finally, perform a Concat operation on Q7 and Q10 to complete the splicing along the channel dimension and obtain the large target cascaded fusion feature map Q11. Perform a Conv1×1 convolution operation on Q11 to obtain the large target cascaded fusion feature map Q12. Q12 is the output feature map of the DCF_L module.

[0055] As a preferred embodiment of the aerial remote sensing image target detection method based on the HSNet model described in this invention, the construction and execution process of the DCF_S module in S32 is as follows:

[0056] S321: Treat the input feature map of the DCF_S module as a small target cascaded fusion feature map Z1, input Z1 into the multi-branch redundant feature generation module for small target feature maps, and output the small target cascaded fusion feature map Z2.

[0057] S322: Z2 is then processed through two branches to further process and fuse target features:

[0058] In the first branch, MaxPool is first performed on Z2, using a 3×3 pooling window that slides with a stride of 1 pixel and fills the edges of Z2 with a full ring of pixels to highlight salient features, resulting in the small object cascaded fusion feature map Z3. Simultaneously, AvgPool is performed on Z2, using a 3×3 pooling window that slides with a stride of 1 pixel and fills the edges of Z2 with a full ring of pixels to suppress noise and smooth features, resulting in the small object cascaded fusion feature map Z4. Then, Z3 and Z4 are concatenated along the channel dimension to obtain the small object cascaded fusion feature map Z5. Next, Z5 is convolved with a Conv1×1 convolution to make the features more compact, resulting in the small object cascaded fusion feature map Z6. Finally, Z6 is input into Channel Attention. The Channel Attention Module learns to obtain the importance weight vector W3 for different channels, which is used to evaluate the importance of each feature channel. Then, based on W3, the channel dimensions of Z6 are weighted, that is, W3 is multiplied by Z6 channel by channel to obtain the small target cascaded fusion feature map Z7.

[0059] In the second branch, Z2 is first subjected to a Conv3×3 convolution operation to further extract and transform the features, resulting in a small target cascaded fusion feature map Z8. Then, Z8 is subjected to a Conv1×1 convolution operation to obtain a small target cascaded fusion feature map Z9. Z9 is then input into the Spatial Attention Module to learn the importance weight matrix W4 of the feature map in the spatial dimension, which is used to evaluate the importance of each spatial location. Then, the spatial locations of Z9 are weighted according to W4, that is, W4 is multiplied with Z9 position by position to obtain a small target cascaded fusion feature map Z10.

[0060] S323: Finally, perform a Concat operation on Z7 and Z10 to complete the splicing along the channel dimension, and obtain the small target cascaded fusion feature map Z11; perform a Conv1×1 convolution operation on Z11 to obtain the small target cascaded fusion feature map Z12; Z12 is the output feature map of the DCF_S module.

[0061] As a preferred embodiment of the aerial remote sensing image target detection method based on the HSNet model described in this invention, the specific steps of S4 are as follows:

[0062] S41: Input any aerial remote sensing target detection image F1 into the multi-scale context feature enhancement extractor in the HSNet model. Through multi-scale context expansion and spatial detail feature enhancement, aerial remote sensing enhanced feature maps P1, P2 and P3 are obtained.

[0063] S42: Input P1, P2 and P3 into the adaptive feature refinement fusion unit in the HSNet model to achieve adaptive fusion and refinement of multi-scale target features, and obtain the refined fusion feature maps R1, R2 and R3 of airborne remote sensing.

[0064] S43: Input R1, R2, and R3 into three independent target detection heads respectively. Each head obtains a prediction tensor containing target detection information from the aerial remote sensing image. Then, merge the prediction results of the three heads together and use the non-maximum suppression algorithm (NMS) to filter high-confidence detection boxes. Finally, output the detection result image with target annotations.

[0065] Compared with existing technologies:

[0066] This invention effectively improves the detection capability of small targets in aerial remote sensing image target detection, while enhancing spatial location awareness and multi-scale feature fusion performance. First, the refined spatial feature capture mechanism of the DHFaster module in the multi-scale contextual feature enhancement extractor, combined with the multi-scale dilated convolution of the MCM module, significantly enhances the feature representation capability of small targets, effectively mitigating the problems of missed and false detections of small targets in complex backgrounds. Second, the DHFaster module in the multi-scale contextual feature enhancement extractor employs a grouped non-shared convolution mechanism, fully utilizing spatial location information to improve the localization capability of irregular targets. Finally, the adaptive feature refinement fusion engine strengthens the feature representation of small and large targets respectively through the DCF_S and DCF_L modules, and combines a bidirectional fusion path to achieve cross-scale adaptive fusion, solving the problems of feature suppression and fusion imbalance caused by large scale differences among multiple targets. The overall model structure is designed specifically for the characteristics of aerial remote sensing images, maintaining high detection accuracy and stability even in complex scenes. Attached Figure Description

[0067] Figure 1 This is a schematic diagram of the process of the present invention;

[0068] Figure 2 This is a structural diagram of the multi-scale context feature enhancement extractor of the present invention;

[0069] Figure 3This is a structural design diagram of the DHFaster module of the present invention;

[0070] Figure 4 This is a structural design diagram of the MCM module of the present invention;

[0071] Figure 5 This is a structural diagram of the adaptive feature refining fusion processor of the present invention;

[0072] Figure 6 This is a schematic diagram of the overall structure of the DCF_L of the present invention;

[0073] Figure 7 This is a structural design diagram of the DCF_S module of the present invention;

[0074] Figure 8 This is a diagram of the overall architecture of the HSNet model of this invention. Detailed Implementation

[0075] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0076] This invention provides a target detection method for aerial remote sensing images based on the HSNet model. Please refer to [link to relevant documentation]. Figure 1-6 The specific steps are as follows:

[0077] S1: Construct a target detection dataset for aerial remote sensing images;

[0078] The specific steps of S1 are as follows:

[0079] S11: Based on existing public datasets, collect aerial remote sensing images and construct an initial target detection dataset;

[0080] Among them, publicly available aerial remote sensing image datasets include HRRSD, DIOR, etc.

[0081] Released by the University of Chinese Academy of Sciences in 2019, HRRSD contains 21,761 remote sensing images from Google Earth and Baidu Maps, with spatial resolutions ranging from 0.15m to 1.2m. It includes 55,740 target instances, with approximately 4,000 instances per target category.

[0082] DIOR, developed by the Matricia laboratory of the French Ministry of Defense, aims to address detection challenges in complex scenarios such as multi-view, occlusion, and deformation. The dataset contains 23,463 remote sensing images and 192,472 axis-aligned bounding box annotations, with all images uniformly sized at 800×800 pixels.

[0083] S12: Preprocess the initial target detection dataset to obtain the preprocessed target detection dataset. The main preprocessing operations include: firstly, cropping the aerial remote sensing images in the initial target detection dataset to a uniform size and performing image channel normalization to eliminate imaging differences; then, data augmentation of the images is performed by adding random noise and simulating cloud and fog occlusion to improve the robustness of the model under complex conditions.

[0084] S13: Divide the preprocessed target detection dataset to obtain the aerial remote sensing image target detection dataset; the specific design of the division process is as follows: divide the preprocessed target detection dataset into training set, validation set and test set according to a scientific ratio to ensure that each type of target is evenly distributed in each subset, while taking into account scene diversity, and providing a structured data foundation for model training, parameter tuning and other tasks.

[0085] Example:

[0086] The DIOR (Digital Intrusion Orbit) public dataset of aerial remote sensing images was selected as the data source. Two hundred thousand 800×800 pixel remote sensing images were chosen from this dataset, all of which already include axis-aligned bounding box annotations. The selected images cover a variety of typical scenes, ensuring good diversity and representativeness in both category and scene distribution, laying the foundation for constructing a high-quality initial object detection dataset. After preprocessing, the preprocessed dataset was divided into training, validation, and test sets in a 7:2:1 ratio. During the partitioning process, it was ensured that each category of targets was evenly distributed across the subsets, while also covering multiple scene types.

[0087] S2: Construct a multi-scale contextual feature enhancement extractor; the structure of the multi-scale contextual feature enhancement extractor is as follows: Figure 2 As shown.

[0088] The specific construction process of the multi-scale contextual feature enhancement extractor in S2 is as follows:

[0089] S21: Input any aerial remote sensing target detection image F1 into the DHFaster module, and output the aerial remote sensing target detection feature map F2 by efficiently capturing refined spatial features and contextual information;

[0090] S22: Then F2 is input into the DHFaster module to further capture refined spatial features and contextual information, and outputs the airborne remote sensing target detection feature map F3;

[0091] S23: Then input F3 into the MCM module to enhance the feature extraction capability of the target and output the aerial remote sensing target detection feature map F4;

[0092] S24: Input F4 into the DHFaster module and output the aerial remote sensing target detection feature map F5. Then input F5 into the MCM module and output the aerial remote sensing enhancement feature map P1.

[0093] S25: Input P1 into the DHFaster module and output the aerial remote sensing target detection feature map F6. Then input F6 into the MCM module and output the aerial remote sensing enhancement feature map P2.

[0094] S26: Input P2 into the DHFaster module and output the aerial remote sensing target detection feature map F7. Then input F7 into the MCM module and output the aerial remote sensing target detection feature map F8.

[0095] S27: Finally, input F8 into the SPPF module to generate the aerial remote sensing enhanced feature map P3;

[0096] S28: After processing by the multi-scale contextual feature enhancement extractor described above, aerial remote sensing enhanced feature maps P1, P2 and P3 are obtained.

[0097] The SPPF module, a core improvement in object detection networks, is the successor to the SPP module. Its core function is to address the limitation of Convolutional Neural Networks (CNNs) requiring a fixed input image size. SPPF performs multi-scale pooling operations on the input feature map using multiple max-pooling kernels of different sizes (e.g., 5x5, 9x9, 13x13) in parallel. It then concatenates these pooling results with the unpooled original feature map, efficiently fusing local details and global contextual information into a fixed-length output vector. This design significantly enhances the model's ability to extract multi-scale features from a single image, dramatically improving detection performance for objects of varying sizes.

[0098] The DHFaster module is a convolutional architecture specifically designed for target detection in aerial remote sensing images. It aims to address the spatial information loss and feature confusion issues caused by shared convolutional parameters in multi-scale target detection. The module employs a dual-branch architecture: the first branch preserves the basic features of the spatial structure through convolutional operations and promotes cross-channel information interaction and fusion; the second branch, based on grouping, divides the input features into multiple non-overlapping local receptive fields, encoding them separately using independent and non-shared convolutional kernels, thereby fully capturing subtle feature differences at different spatial locations and enhancing spatial perception. Finally, by fusing the outputs of the two branches, DHFaster effectively integrates local and global spatial context information, significantly improving the model's ability to represent features of targets with diverse scales and complex backgrounds in aerial images. The specific structural design of the DHFaster module is as follows: Figure 3 As shown.

[0099] The construction and execution process of the DHFaster module in S21 is as follows:

[0100] S211: Treat the input feature map of the DHFaster module as a context-interactive fusion feature map H1, and process H1 through two branches:

[0101] In the first branch, H1 is first subjected to a Conv1×1 convolution operation to initially adjust the number of channels and retain the original spatial location information, resulting in a context-interactive fusion feature map H2; then, H2 is subjected to a Conv3×3 convolution operation to extract information on local image features, resulting in a context-interactive fusion feature map H3; finally, H3 is subjected to a Conv1×1 convolution operation to fuse features from different channels at the same location, resulting in a context-interactive fusion feature map H4.

[0102] Wherein: Conv3×3 convolution operation represents the process of processing the feature image through a convolution kernel of size 3×3; Conv1×1 convolution operation represents the process of processing the feature image through a convolution kernel of size 1×1.

[0103] In the second branch, H1 is first divided into three groups—Group1, Group2, and Group3—along the channel dimension, allowing different groups to focus on different types of feature patterns. Then, the feature map of each group is further divided into k×k non-overlapping local sub-regions. Each local sub-region uses an independent, non-shared convolutional kernel for feature extraction, ensuring the model can capture and preserve subtle feature differences at each unique location. Afterward, all processed, size-invariant sub-regions are reassembled in their original positions, restoring the feature map to its original size before sub-region division. After these operations, Group1... Group 1 yields context interaction fusion feature map H5_1, Group 2 yields context interaction fusion feature map H5_2, and Group 3 yields context interaction fusion feature map H5_3. Then, H5_1, H5_2, and H5_3 are concatenated along the channel dimension to generate context interaction fusion feature map H5. Next, H5 is convolved with a Conv1×1 convolution to achieve information exchange and fusion among the three groups, resulting in context interaction fusion feature map H6. Finally, H6 is convolved with a Conv3×3 convolution to generate context interaction fusion feature map H7.

[0104] S212: Finally, perform a Concat operation on H4 and H7 to complete the splicing along the channel dimension and generate a contextual interactive fusion feature map H8; H8 is the output feature map of the DHFaster module.

[0105] The MCM module is a feature enhancement module specifically designed to address the challenge of small target detection in aerial remote sensing images. This module optimizes feature richness and expands the receptive field through a multi-branch structure and dilated convolution techniques. Its core structure comprises three parallel branches, each employing dilated convolutions with different dilation rates for feature extraction, simultaneously capturing local details, mid-range context, and global macroscopic features. Each branch combines residual connections to preserve the original feature information, which is then integrated via an adaptive weighted fusion mechanism to dynamically balance the contributions of multi-scale features. Finally, residual connections add the fused features to the original input, significantly increasing the effective receptive field while highlighting discriminative semantic features and suppressing background interference. The MCM module effectively alleviates problems such as high target feature similarity, weak representation of small targets, and feature overload caused by dense targets, significantly improving the detection accuracy and robustness of multi-scale targets in complex backgrounds. The specific structural design of the MCM module is as follows: Figure 4 As shown.

[0106] The construction and execution process of the MCM module in S23 is as follows:

[0107] S231: Treat the input feature map of the MCM module as a multi-domain enhanced fusion feature map X1, and process X1 through three branches:

[0108] In the first branch, X1 is first subjected to a Conv1×1 convolution operation to obtain a multi-domain enhanced fusion feature map X2; then X2 is subjected to a Conv3×3 convolution operation to extract local spatial features to obtain a multi-domain enhanced fusion feature map X3; then X3 is input into a dilated convolution module DilatedConv with a kernel size of 3×3 and a dilation rate of 1 to further enhance and consolidate the local feature expression to obtain a multi-domain enhanced fusion feature map X4; X1 and X4 are subjected to a Concat operation to complete the concatenation along the channel dimension to enrich the hierarchy of features and generate a multi-domain enhanced fusion feature map X5;

[0109] Dilated convolution is a convolutional operation that expands the receptive field by inserting holes between kernel elements, enabling it to capture a wider range of contextual information while keeping the number of parameters constant. Compared to traditional convolution, dilated convolution controls the sampling interval by adjusting the dilation rate, making it suitable for tasks requiring multi-scale feature extraction.

[0110] In the second branch, X1 is first subjected to a Conv1×1 convolution operation to obtain a multi-domain enhanced fusion feature map X6; then X6 is subjected to a Conv3×3 convolution operation to extract local spatial features, resulting in a multi-domain enhanced fusion feature map X7; X7 is then input into a dilated convolution module DilatedConv with a kernel size of 3×3 and a dilation rate of 3 to capture mid-range contextual information, resulting in a multi-domain enhanced fusion feature map X8; X1 and X8 are then subjected to a Concat operation to complete the concatenation along the channel dimension to enrich the hierarchical nature of the features, generating a multi-domain enhanced fusion feature map X9;

[0111] In the third branch, X1 is first subjected to a Conv1×1 convolution operation to obtain a multi-domain enhanced fusion feature map X10; then, X10 is subjected to a Conv3×3 convolution operation to extract local spatial features, resulting in a multi-domain enhanced fusion feature map X11; then, X11 is input into a DilatedConv convolution module with a kernel size of 3×3 and a dilation rate of 5 to capture long-range dependencies and global scene context, resulting in a multi-domain enhanced fusion feature map X12; finally, X1 and X12 are subjected to a Concat operation to complete the concatenation along the channel dimension to enrich the hierarchy of features, generating a multi-domain enhanced fusion feature map X13;

[0112] S232: Subsequently, through three learnable fusion weights Adaptive fusion is performed on X5, X9, and X13 to generate a multi-domain enhanced fusion feature map X14; the formula for adaptive fusion is as follows:

[0113]

[0114] in, It represents the horizontal position of a pixel on the feature map; It represents the vertical position of a pixel on the feature map; This indicates the number of a specific channel on the feature map. This indicates that in X5, it is located at the... Column, No. line, number The value at that specific location in each channel; This indicates that in X9, the position is at the... Column, No. line, number The value at that specific location in each channel; This indicates that in X13, the position located at the... Column, No. line, number The value at that specific location in each channel.

[0115] S233: Then, perform a Conv1×1 convolution operation on X14 to adjust the number of channels of X14 to match the number of channels of X1, and obtain the multi-domain enhanced fusion feature map X15; finally, add X1 and X15 element by element to prevent the gradient from vanishing in the deep network, and obtain the multi-domain enhanced fusion feature map X16; X16 is the output feature map of the MCM module.

[0116] S3: Construct an adaptive feature refinement fusion engine; the structure of the adaptive feature refinement fusion engine is as follows: Figure 5 As shown.

[0117] The specific construction process of the adaptive feature refinement fusion unit in S3 is as follows:

[0118] S31: First, input the enhanced feature map P3 of the airborne remote sensing obtained by the multi-scale context feature enhancement extractor into the DCF_L module to enhance the representation ability of large targets and overall structure, and output the airborne remote sensing target detection feature map F9.

[0119] S32: Then, upsample F9 to generate the aerial remote sensing target detection feature map F10; then, add the aerial remote sensing enhanced feature map P2 obtained by the multi-scale context feature enhancement extractor and F10 element by element to obtain the aerial remote sensing target detection feature map F11; input F11 into the C2f module (n=3) to obtain the aerial remote sensing target detection feature map F12; input F12 into the DCF_S module to further enhance the representation ability of small targets in the features, and output the aerial remote sensing target detection feature map F13; then, upsample F13 to generate the aerial remote sensing target detection feature map F14;

[0120] S33: Next, the aerial remote sensing enhanced feature maps P1 and F14 obtained by the multi-scale context feature enhancement extractor are added element by element to obtain the aerial remote sensing target detection feature map F15; F15 is input into the C2f module (n=3) to refine the target features and obtain the aerial remote sensing refined fusion feature map R1.

[0121] S34: Input R1 into the DHFaster module to further capture refined spatial features and contextual information, and output the aerial remote sensing target detection feature map F16; then add F13 and F16 element by element to obtain the aerial remote sensing target detection feature map F17.

[0122] S35: Input F17 into the C2f module (n=3) to further refine the target features and obtain the refined fusion feature map R2 of airborne remote sensing; input R2 into the DHFaster module to deeply capture refined spatial features and contextual information and output the airborne remote sensing target detection feature map F18.

[0123] S36: Then add F9 and F18 element by element to obtain the aerial remote sensing target detection feature map F19; then input F19 into the C2f module (n=3) to perform deep refinement of the target features and obtain the aerial remote sensing refined fusion feature map R3.

[0124] S37: After processing by the above adaptive feature refinement fusion processor, the refined and fused feature maps R1, R2 and R3 of airborne remote sensing are obtained.

[0125] The C2f module is a highly efficient cross-stage feature fusion module, whose core consists of n bottleneck structures. This module divides the input features into two parts through a split operation. One part undergoes rich feature transformation and extraction through multiple bottlenecks, while the other part retains the original feature information. Finally, the transformed features are fused with the original features through concatenation and convolution operations.

[0126] This paper innovatively designs the DCF_L and DCF_S modules in an adaptive feature refinement fusion processor to address the fusion challenge caused by significant differences in target scale in aerial remote sensing image target detection. Traditional methods often suffer from large targets suppressing smaller targets or loss of detail in small targets due to downsampling operations when fusing multi-scale features, severely impacting the accuracy and robustness of multi-scale target detection. To address this, two modules are specifically designed: the DCF_L module focuses on large targets and overall structural representation, utilizing a larger receptive field to capture the global context and contour information of the target, and suppressing background interference through feature calibration to enhance the feature response of large target regions; the DCF_S module focuses on enhancing small target features, significantly increasing the weight of small target regions during fusion by strengthening local detail extraction and high-frequency information preservation, and combining a dynamic weighting mechanism, effectively preventing the features of small targets from being submerged during the fusion process. The two work together to balance the preservation and fusion of features of targets of different sizes. DCF_L ensures the accurate positioning and shape integrity of large targets, while DCF_S ensures the detection rate and detail integrity of small targets. Together, they achieve adaptive and balanced fusion of multi-scale features, significantly improving the overall performance of aerial image target detection in complex scenes.

[0127] The DCF_L module is specifically designed for large target detection in aerial remote sensing images. This module captures the global structural information and contextual semantics of the target by constructing a multi-branch structure, effectively perceiving the overall contour and spatial distribution of large targets. Simultaneously, it utilizes a dual-attention mechanism to adaptively calibrate channel and spatial weights, effectively focusing on salient regions of large targets. The DCF_L module significantly improves the accuracy of large target localization and shape integrity detection in complex scenes, overcoming the structural information loss and fusion imbalance problems caused by scale differences in traditional methods. The specific structural design of the DCF_L module is as follows: Figure 6 As shown.

[0128] The construction and execution process of the DCF_L module in S31 is as follows:

[0129] S311: Treat the input feature map of the DCF_L module as a large target cascaded fusion feature map Q1, and input Q1 into the multi-branch redundant feature generation module for the large target feature map to obtain the large target cascaded fusion feature map Q2. The specific process of obtaining Q2 through the multi-branch redundant feature generation module for the large target feature map is as follows:

[0130] A Conv5×5 convolution operation is performed on Q1 to capture the local detailed features of the large target, resulting in a redundant generated feature map L1. Then, BN batch normalization is performed on L1 to obtain a redundant generated feature map L2. Finally, the SiLU activation function is used to process L2 to obtain a redundant generated feature map L3.

[0131] Meanwhile, a DWConv5×5 depthwise separable convolution operation is performed on Q1 to expand the receptive field while reducing the computational cost, resulting in a large target redundancy generated feature map L4. Subsequently, BN batch normalization is performed on L4 to obtain a large target redundancy generated feature map L5. Finally, the SiLU activation function is used to process L5 to obtain a large target redundancy generated feature map L6.

[0132] Among them, depthwise separable convolution is a lightweight convolution operation that decomposes standard convolution into two steps: depthwise convolution and pointwise convolution, which can significantly reduce the amount of computation and parameters.

[0133] Next, Q1, L3, and L6 are concatenated along the channel dimension to construct a comprehensive and redundant composite feature map, resulting in the large target redundant generated feature map L7. Then, L7 is convolved using a Conv1×1 method to fuse and reduce the dimensionality of the concatenated feature maps, yielding the large target cascaded fused feature map Q2.

[0134] S312: Q2 is then processed through two branches to further process and fuse the target features:

[0135] In the first branch, MaxPool is first performed on Q2, using a 3×3 pooling window that slides with a stride of 1 pixel and fills the edges of Q2 with a full ring of pixels to highlight salient features, resulting in the large target cascaded fusion feature map Q3. Simultaneously, AvgPool is performed on Q2, using a 3×3 pooling window that slides with a stride of 1 pixel and fills the edges of Q2 with a full ring of pixels to suppress noise and smooth features, resulting in the large target cascaded fusion feature map Q4. Then, Q3 and Q4 are concatenated along the channel dimension to obtain the large target cascaded fusion feature map Q5. Finally, a Conv1×1 convolution operation is performed on Q5 to make the features more compact, resulting in the large target cascaded fusion feature map Q6. Q6 is then input into Channel Attention. The Channel Attention Module learns to obtain importance weight vectors W1 for different channels, which are used to evaluate the importance of each feature channel. Then, based on W1, a weighting operation is performed on the channel dimensions of Q6, that is, W1 is multiplied by Q6 channel by channel to obtain the large target cascaded fusion feature map Q7.

[0136] The Channel Attention Module aims to enhance key channel features and suppress irrelevant channel information by learning the importance weights of each channel in the feature map. After receiving the input feature map, this module first aggregates its global information, obtaining channel statistics from different dimensions through global average pooling. Then, it processes the aggregated information using a shared Multilayer Perceptron (MLP) to generate a weight vector matching the number of input channels.

[0137] In the second branch, Q2 is first subjected to a Conv3×3 convolution operation to further extract and transform the features, resulting in a large target cascaded fusion feature map Q8. Then, Q8 is subjected to a Conv1×1 convolution operation to obtain a large target cascaded fusion feature map Q9. Next, Q9 is input into the Spatial Attention Module to learn the importance weight matrix W2 of the feature map in the spatial dimension, which is used to evaluate the importance of each spatial position. Then, the spatial positions of Q9 are weighted according to W2, that is, W2 is multiplied with Q9 position by position to obtain a large target cascaded fusion feature map Q10.

[0138] The Spatial Attention Module focuses on enhancing information about key spatial locations in the feature map by learning the importance weights of the spatial dimensions to highlight the target region. After receiving the input feature map, this module first performs max pooling and average pooling operations on it along the channel dimension to obtain two different spatial feature descriptions. These two feature maps are then concatenated and convolutionally processed to generate a weight matrix with the same size as the input space.

[0139] S313: Finally, perform a Concat operation on Q7 and Q10 to complete the splicing along the channel dimension and obtain the large target cascaded fusion feature map Q11. Perform a Conv1×1 convolution operation on Q11 to obtain the large target cascaded fusion feature map Q12. Q12 is the output feature map of the DCF_L module.

[0140] The DCF_S module is specifically designed to enhance the performance of small target detection in remote sensing images. This module focuses on extracting subtle local features by constructing a multi-branch structure and utilizes a dual-attention mechanism to adaptively calibrate channel and spatial weights, effectively focusing on the salient regions of small targets. The DCF_S module significantly enhances the feature representation capability of small targets in complex scenes, effectively improving the detection rate and localization accuracy of small targets, providing crucial technical support for multi-scale target detection systems. The specific structural design of the DCF_S module is as follows: Figure 7 As shown.

[0141] The construction and execution process of the DCF_S module in S32 is as follows:

[0142] S321: The input feature map of the DCF_S module is regarded as a small target cascaded fusion feature map Z1. Z1 is input into the multi-branch redundant feature generation module for small target feature maps, and the output is the small target cascaded fusion feature map Z2. The specific process of obtaining Z2 through the multi-branch redundant feature generation module for small target feature maps is as follows:

[0143] A Conv3×3 convolution operation is performed on Z1 to capture the local detailed features of the small target, resulting in a redundant generated feature map S1. Then, BN batch normalization is performed on S1 to obtain a redundant generated feature map S2. Finally, the SiLU liveness function is used to process S2 to obtain a redundant generated feature map S3.

[0144] Meanwhile, Z1 is input into a dilated convolutional module DilatedConv with a kernel size of 3×3 and a dilation rate of 2 to effectively capture the contextual information around the small target, resulting in a small target redundant generation feature map S4; then S4 is input into a dilated convolutional module DilatedConv with a kernel size of 3×3 and a dilation rate of 4 to further capture the surrounding contextual information, resulting in a small target redundant generation feature map S5;

[0145] Next, Z1, S3, and S5 are concatenated along the channel dimension to construct a comprehensive and redundant composite feature map, resulting in a small target redundant generation feature map S6. Then, S6 is convolved using a Conv1×1 operation to fuse the concatenated feature maps, yielding a small target cascaded fused feature map Z2.

[0146] S322: Z2 is then processed through two branches to further process and fuse target features:

[0147] In the first branch, MaxPool is first performed on Z2, using a 3×3 pooling window that slides with a stride of 1 pixel and fills the edges of Z2 with a full ring of pixels to highlight salient features, resulting in the small object cascaded fusion feature map Z3. Simultaneously, AvgPool is performed on Z2, using a 3×3 pooling window that slides with a stride of 1 pixel and fills the edges of Z2 with a full ring of pixels to suppress noise and smooth features, resulting in the small object cascaded fusion feature map Z4. Then, Z3 and Z4 are concatenated along the channel dimension to obtain the small object cascaded fusion feature map Z5. Next, Z5 is convolved with a Conv1×1 convolution to make the features more compact, resulting in the small object cascaded fusion feature map Z6. Finally, Z6 is input into Channel Attention. The Channel Attention Module learns to obtain the importance weight vector W3 for different channels, which is used to evaluate the importance of each feature channel. Then, based on W3, the channel dimensions of Z6 are weighted, that is, W3 is multiplied by Z6 channel by channel to obtain the small target cascaded fusion feature map Z7.

[0148] In the second branch, Z2 is first subjected to a Conv3×3 convolution operation to further extract and transform the features, resulting in a small target cascaded fusion feature map Z8. Then, Z8 is subjected to a Conv1×1 convolution operation to obtain a small target cascaded fusion feature map Z9. Z9 is then input into the Spatial Attention Module to learn the importance weight matrix W4 of the feature map in the spatial dimension, which is used to evaluate the importance of each spatial location. Then, the spatial locations of Z9 are weighted according to W4, that is, W4 is multiplied with Z9 position by position to obtain a small target cascaded fusion feature map Z10.

[0149] S323: Finally, perform a Concat operation on Z7 and Z10 to complete the splicing along the channel dimension, and obtain the small target cascaded fusion feature map Z11; perform a Conv1×1 convolution operation on Z11 to obtain the small target cascaded fusion feature map Z12; Z12 is the output feature map of the DCF_S module.

[0150] S4: Based on a multi-scale contextual feature enhancement extractor and an adaptive feature refinement fusion mechanism, an HSNet model is constructed for target detection in aerial remote sensing images; the overall architecture of the HSNet model is as follows. Figure 8 As shown.

[0151] The specific steps of S4 are as follows:

[0152] S41: Input any aerial remote sensing target detection image F1 into the multi-scale contextual feature enhancement extractor in the HSNet model. Through multi-scale contextual expansion and spatial detail feature enhancement, aerial remote sensing enhanced feature maps P1, P2, and P3 are obtained; the specific process is as follows:

[0153] First, the aerial remote sensing target detection image F1 is input into the DHFaster module. By efficiently capturing refined spatial features and contextual information, the aerial remote sensing target detection feature map F2 is output. Then, F2 is input into the DHFaster module to further capture refined spatial features and contextual information, resulting in the aerial remote sensing target detection feature map F3. Finally, F3 is input into the MCM module to enhance the target feature extraction capability, outputting the aerial remote sensing target detection feature map F4. Inputting F4 into the DHFaster module outputs the aerial remote sensing target detection feature map F5; then inputting F5 into the MCM module outputs the aerial remote sensing enhanced feature map P1; inputting P1 into the DHFaster module outputs the aerial remote sensing target detection feature map F6; then inputting F6 into the MCM module outputs the aerial remote sensing enhanced feature map P2; inputting P2 into the DHFaster module outputs the aerial remote sensing target detection feature map F7; then inputting F7 into the MCM module outputs the aerial remote sensing target detection feature map F8; finally, inputting F8 into the SPPF module generates the aerial remote sensing enhanced feature map P3.

[0154] S42: Input P1, P2, and P3 into the adaptive feature refinement fusion unit in the HSNet model to achieve adaptive fusion and refinement of multi-scale target features, obtaining refined fused feature maps R1, R2, and R3 from airborne remote sensing; the specific process is as follows:

[0155] First, the enhanced airborne remote sensing feature map P3 obtained from the multi-scale contextual feature enhancement extractor is input into the DCF_L module to strengthen the representation of large targets and overall structures, outputting the airborne remote sensing target detection feature map F9. Then, F9 is upsampled to generate the airborne remote sensing target detection feature map F10. Next, the enhanced airborne remote sensing feature maps P2 and F10 are added element-wise to obtain the airborne remote sensing target detection feature map F11. F11 is input into the C2f module (n=3) to obtain the airborne remote sensing target detection feature map F12. F12 is then input into the DCF_S module to further enhance the representation of small targets, outputting the airborne remote sensing target detection feature map F13. Next, F13 is upsampled to generate the airborne remote sensing target detection feature map F14. Finally, the enhanced airborne remote sensing feature maps P1 and F14 are added element-wise to obtain the airborne remote sensing target detection feature map F15. F15 is input into the C2f module (n=3) to refine the target features, resulting in the refined fusion feature map R1 of the airborne remote sensing. R1 is then input into the DHFaster module to further capture refined spatial features and contextual information, outputting the airborne remote sensing target detection feature map F16. F13 and F16 are then added element-wise to obtain the airborne remote sensing target detection feature map F17. F17 is input into the C2f module (n=3) to further refine the target features, resulting in the refined fusion feature map R2 of the airborne remote sensing. R2 is input into the DHFaster module to deeply capture refined spatial features and contextual information, outputting the airborne remote sensing target detection feature map F18. F9 and F18 are then added element-wise to obtain the airborne remote sensing target detection feature map F19. Finally, F19 is input into the C2f module (n=3) to perform deep refinement of the target features, resulting in the refined fusion feature map R3 of the airborne remote sensing.

[0156] S43: Input R1, R2, and R3 into three independent target detection heads, each of which produces a prediction tensor containing target detection information from the aerial remote sensing image. In the prediction tensor, each row corresponds to a prediction result, which includes the bounding box coordinates of the target in the aerial remote sensing image, the target's corresponding category label, and a confidence score. The target's bounding box is determined by the coordinates of the top-left and bottom-right pixels. Subsequently, the prediction results of the three heads are merged together, and high-confidence detection boxes are filtered using the Non-Maximum Suppression (NMS) algorithm. Finally, a detection result image with target annotations is output. The target annotations specifically include: the target's corresponding category, the coordinates of the target's bounding box, and the confidence score.

[0157] The target prediction head is the core output component that has been publicly disclosed. It is responsible for directly predicting the bounding box, class, and confidence score of the target. It typically consists of a series of convolutional layers that transform deep feature maps into dense prediction tensors, where each grid cell predicts multiple anchor boxes and outputs the box's position offset, class probability, and confidence score indicating the presence of the target.

[0158] Non-Maximum Suppression (NMS) is a post-processing algorithm for object detection, primarily used to eliminate redundant detection boxes. Its core idea is to retain the prediction box with the highest confidence while suppressing highly overlapping neighboring boxes with lower confidence, thus ensuring that only the most accurate detection result is output for each object.

[0159] Example:

[0160] Let the dimensions of each of the following images or feature maps be H×W×C, where H is the height of the feature map, W is the width of the feature map, and C is the number of channels of the feature map.

[0161] An aerial remote sensing target detection image with a size of 800×800×3 is input into a multi-scale contextual feature enhancement extractor, and this aerial remote sensing target detection image is denoted as aerial remote sensing target detection image F1.

[0162] First, F1 is input into the first DHFaster module. In the DHFaster module, F1 is used as input to the context interaction fusion feature map H1. In the first branch, H1 undergoes a Conv1×1 convolution operation to initially adjust the number of channels, generating a context interaction fusion feature map H2 with a size of 800×800×32. Subsequently, H2 undergoes a Conv3×3 convolution operation for spatial downsampling, generating a context interaction fusion feature map H3 with a size of 400×400×64. Finally, a Conv1×1 convolution operation is performed on H3 to enhance channel interaction, generating a context interaction fusion feature map H4 with a size of 400×400×64. In the second branch, H1 is divided into three groups along the channel dimension. The feature map of each group is divided into non-overlapping local sub-regions, which are then processed and recombined using independent convolutional kernels. The outputs of the three groups are then concatenated to generate a context-interactive fusion feature map H5 with a size of 800×800×3. A Conv1×1 convolution is then performed on H5 to achieve inter-group information interaction, generating a context-interactive fusion feature map H6 with a size of 800×800×32. Finally, a Conv3×3 convolution operation is performed on H6 for spatial downsampling, generating a context-interactive fusion feature map H7 with a size of 400×400×64. H4 and H7 are then concatenated to finally output a context-interactive fusion feature map H8 with a size of 400×400×128, which is the aerial remote sensing target detection feature map F2.

[0163] Subsequently, F2 is input as the context interaction fusion feature map H1 into the second DHFaster module. After the same processing flow, the airborne remote sensing target detection feature map F3 with a size of 200×200×256 is output.

[0164] F3 is input into the first MCM module. In the MCM module, F3 is used as input to the multi-domain enhanced fusion feature map X1. In the first branch, X1 is processed by Conv1×1 to adjust the channels, generating a multi-domain enhanced fusion feature map X2 with a size of 200×200×256; X2 is processed by Conv3×3 to extract local features, generating a multi-domain enhanced fusion feature map X3 with a size of 200×200×256; X3 is then input into a dilated convolution module with a kernel size of 3×3, a dilation rate of 1, and padding=1, generating a multi-domain enhanced fusion feature map X4 with a size of 200×200×256; X1 and X4 are concatenated to generate a multi-domain enhanced fusion feature map X5 with a size of 200×200×512. In the second branch, X1 is adjusted by Conv1×1 to become a multi-domain enhanced fusion feature map X6 with a size of 200×200×256; X6 is converted into a multi-domain enhanced fusion feature map X7 with a size of 200×200×256 by Conv3×3 convolution operation; X7 is input into a dilated convolution module with a kernel size of 3×3, an expansion rate of 3, and padding=3 to generate a multi-domain enhanced fusion feature map X8 with a size of 200×200×256; X1 and X8 are concatenated to generate X9 with a size of 200×200×512. In the third branch, X1 is adjusted by a Conv1×1 convolution operation to become a multi-domain enhanced fusion feature map X10 with a size of 200×200×256; X10 is then converted by a Conv3×3 convolution operation to generate a multi-domain enhanced fusion feature map X11 with a size of 200×200×256; X11 is input into a dilated convolution module with a kernel size of 3×3, an expansion rate of 5, and padding of 5 to generate a multi-domain enhanced fusion feature map X12 with a size of 200×200×256; X1 and X12 are concatenated to generate X13 with a size of 200×200×512.

[0165] Through three learnable fusion weights Adaptive fusion of X5, X9, and X13 is performed to generate feature map X14 with a size of 200×200×512. X14 is then subjected to Conv1×1 convolution to obtain X15 with a size of 200×200×256. Finally, X1 and X15 are added element by element to obtain multi-domain enhanced fusion feature map X16 with a size of 200×200×256, which is the aerial remote sensing target detection feature map F4.

[0166] F4 is input into the third DHFaster module. After processing in the same manner as the first two DHFaster modules, an airborne remote sensing target detection feature map F5 with a size of 100×100×512 is output. Then, F5 is input into the second MCM module. After processing in the same manner as the first MCM module, a multi-branch dilated convolution, adaptive fusion and residual addition are performed, resulting in an airborne remote sensing enhancement feature map P1 with a size of 100×100×512.

[0167] Input P1 into the fourth DHFaster module to output an aerial remote sensing target detection feature map F6 with a size of 50×50×1024; then input F6 into the third MCM module to output an aerial remote sensing enhancement feature map P2 with a size of 50×50×1024.

[0168] P2 is input into the fifth DHFaster module, which outputs an aerial remote sensing target detection feature map F7 with a size of 25×25×2048. Then, F7 is input into the fourth MCM module, which outputs an aerial remote sensing target detection feature map F8 with a size of 25×25×2048. Finally, F8 is input into the SPPF module, and through multi-scale pooling and feature concatenation, a final aerial remote sensing enhanced feature map P3 with a size of 25×25×2048 is generated, which integrates multi-scale context.

[0169] The outputs P1, P2, and P3 from the multi-scale context feature enhancement extractor are input into the adaptive feature refinement fusion unit. First, P3 is input into the DCF_L module, serving as the input to the large target cascaded fusion feature map Q1. In the multi-branch redundant feature generation module, Q1 is processed by Conv5×5, BN, and SiLU to generate L3 with a size of 25×25×512. Simultaneously, Q1 is processed by DWConv5×5, BN, and SiLU to generate L6 with a size of 25×25×512. Q1, L3, and L6 are concatenated and then dimensionality-reduced by Conv1×1 to generate the large target cascaded fusion feature map Q2 with a size of 25×25×1024. Q2 is then processed in two branches: In the first branch, max pooling is performed on Q2 using a 3×3 pooling window with a stride of 1 pixel and padding the edges of Q2 with one ring of pixels, resulting in a large target cascaded fusion feature map Q3 with a size of 25×25×1024. Simultaneously, average pooling is performed on Q2 using a 3×3 pooling window with a stride of 1 pixel and padding the edges of Q2 with one ring of pixels, resulting in a large target cascaded fusion feature map Q4 with a size of 25×25×1024. Then, Q3 and Q4 are concatenated along the channel dimension to obtain a large target cascaded fusion feature map Q5 with a size of 25×25×2048. Finally, a Conv1×1 convolution operation is performed on Q5 to obtain a large target cascaded fusion feature map Q6 with a size of 25×25×1024. Q6 is then input into the ChannelAttention Module to learn the importance weight vector W3 for different channels. W3 is then multiplied with Q6 channel-by-channel to obtain a large target cascaded fusion feature map Q7 with a size of 25×25×1024. In the second branch, Q2 is first subjected to a Conv3×3 convolution to obtain a large target cascaded fusion feature map Q8 with a size of 25×25×1024. Q8 is then subjected to a Conv1×1 convolution to obtain a large target cascaded fusion feature map Q9 with a size of 25×25×1024. Q9 is then input into the Spatial Attention Module to learn the spatial importance weight matrix W4 for the feature map. W4 is then multiplied with Q9 position-by-position to obtain a large target cascaded fusion feature map Q10 with a size of 25×25×1024. Finally, Q7 and Q10 are concatenated along the channel dimension to obtain a large target cascaded fusion feature map Q11 with a size of 25×25×2048. A Conv1×1 convolution operation is then performed on Q11 to obtain a large target cascaded fusion feature map Q12 with a size of 25×25×1024. This is the aerial remote sensing target detection feature map F9.

[0170] Upsample F9 by 2 times to generate F10 with a size of 50×50×1024; add F10 to P2 element by element to generate F11 with a size of 50×50×1024; input F11 to the C2f module and output F12 with a size of 50×50×1024.

[0171] F12 is input into the DCF_S module, where it serves as the input to the small target cascaded fusion feature map Z1. In the multi-branch redundant feature generation module, Z1 is processed by Conv3×3, BN, and SiLU to generate S3 with a size of 50×50×256. Simultaneously, Z1 is processed by 3×3 dilated convolution with dilation rates of 2 and 4 to generate S5 with a size of 50×50×256. Z1, S3, and S5 are concatenated and then dimensionality reduced by Conv1×1 to generate Z2 with a size of 50×50×512. Z2 is then processed in two branches: In the first branch, a 3×3 pooling window is first applied to Z2 with a stride of 1 pixel, and a max pooling operation is performed, filling the edges of Z2 with one ring of pixels, resulting in a small target cascaded fusion feature map Z3 with a size of 50×50×512. Simultaneously, a 3×3 pooling window is applied to Z2 with a stride of 1 pixel, and an average pooling operation is performed, filling the edges of Z2 with one ring of pixels, resulting in a small target cascaded fusion feature map Z4 with a size of 50×50×512. Then, Z3 and Z4 are concatenated along the channel dimension to obtain a small target cascaded fusion feature map Z5 with a size of 50×50×1024. Finally, a Conv1×1 convolution operation is performed on Z5 to obtain a small target cascaded fusion feature map Z6 with a size of 50×50×512. Z6 is then input into the Channel Attention Module to learn the importance weight vector W1 for different channels. W1 is then multiplied with Z6 channel-by-channel to obtain a small target cascaded fusion feature map Z7 with a size of 50×50×512. In the second branch, Z2 is first subjected to a Conv3×3 convolution to obtain a small target cascaded fusion feature map Z8 with a size of 50×50×512. Z8 is then subjected to a Conv1×1 convolution to obtain a small target cascaded fusion feature map Z9 with a size of 50×50×512. Z9 is then input into the Spatial Attention Module to learn the spatial importance weight matrix W2 of the feature map. W2 is then multiplied with Z9 spatially to obtain a small target cascaded fusion feature map Z10 with a size of 50×50×512. Finally, Z7 and Z10 are concatenated along the channel dimension to obtain a small target cascaded fusion feature map Z11 with a size of 50×50×1024. A Conv1×1 convolution operation is then performed on Z11 to obtain a small target cascaded fusion feature map Z12 with a size of 50×50×512. Z12 is the aerial remote sensing target detection feature map F13.

[0172] Upsample F13 by a factor of 2 to generate F14 with a size of 100×100×512; add F14 to P1 element-wise to generate F15 with a size of 100×100×512; input F15 to the C2f module to output the refined fused feature map R1 with a size of 100×100×512. Input R1 to the DHFaster module to output F16 with a size of 50×50×512; add F16 to F13 element-wise to generate F17 with a size of 50×50×512; input F17 to the C2f module to output the refined fused feature map R2 with a size of 50×50×512. Input R2 into the DHFaster module, which outputs F18 with a size of 25×25×1024; add F18 and F9 element by element to generate F19 with a size of 25×25×1024; input F19 into the C2f module, which outputs R3, a refined fusion feature map of airborne remote sensing with a size of 25×25×1024.

[0173] After the above processing, the adaptive feature refinement fusion unit outputs refined feature maps at three scales: R1, R2, and R3, which are rich in target feature information at different scales, providing multi-level and multi-scale feature representations for subsequent detection heads.

[0174] Finally, R1, R2, and R3 are input into three independent target detection heads, the prediction results of the three heads are merged together, and the high-confidence detection boxes are filtered by the non-maximum suppression algorithm (NMS). The final output is a detection result image with target annotations, which specifically include: the target's corresponding category, the coordinates of the target bounding box, and the confidence score.

[0175] S5: Based on the aerial remote sensing image target detection dataset, the HSNet model is trained to obtain the optimized HSNet model; the specific process is as follows:

[0176] Before training the HSNet model, all network weight parameters need to be initialized, and various hyperparameters related to the airborne remote sensing target detection task need to be configured.

[0177] After setting the hyperparameters, a batch iterative strategy was used for model training. Each training batch input a certain number of aerial remote sensing images and their annotations into the HSNet model, sequentially passing through a multi-scale contextual feature enhancement extractor, an adaptive feature refinement fusion unit, and the forward propagation of the detection head to generate prediction results. Subsequently, the error between the predicted bounding boxes and the ground truth labels was calculated using loss functions (including classification loss Focal Loss and regression loss CIoU Loss), and the model parameters were dynamically updated using the backpropagation algorithm to optimize network performance. During training, the validation set was used to monitor model performance in real time. Changes in the average accuracy and loss value on the validation set were used to determine if overfitting had occurred, and the learning rate could be dynamically adjusted or an early stopping mechanism could be enabled if necessary. The test set was used for the final evaluation of model performance.

[0178] The model training process is terminated when the validation set loss no longer decreases significantly or the performance stabilizes after multiple consecutive training epochs. Finally, the optimal model weights and parameters from the validation set are saved, resulting in the trained and optimized HSNet aerial remote sensing image target detection model. This model possesses high-precision detection capabilities for multi-scale, multi-class aerial targets, providing reliable support for subsequent practical applications.

[0179] S6: The trained and optimized HSNet model is applied to the target detection scenario of aerial remote sensing images to achieve intelligent identification and localization of targets. Finally, the detection result image is obtained, which includes the target's corresponding category, the coordinates of the target bounding box, and the confidence score, providing reliable technical support and decision-making basis for remote sensing image analysis and other fields.

[0180] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A target detection method for aerial remote sensing images based on the HSNet model, characterized in that, The specific steps are as follows: S1: Construct a target detection dataset for aerial remote sensing images; S2: Construct a multi-scale contextual feature enhancement extractor; S3: Construct an adaptive feature refinement fusion engine; S4: Based on a multi-scale contextual feature enhancement extractor and an adaptive feature refinement fusioner, an HSNet model is constructed for target detection in aerial remote sensing images; S5: Based on the target detection dataset of aerial remote sensing images, the HSNet model is trained to obtain the trained and optimized HSNet model; S6: The trained and optimized HSNet model is applied to the target detection scenario of aerial remote sensing images to achieve intelligent identification and localization of targets. Finally, the detection result image containing the target's corresponding category, the coordinates of the target bounding box, and the confidence score is obtained, providing reliable technical support and decision-making basis for the field of remote sensing image analysis. The specific construction process of the multi-scale contextual feature enhancement extractor in S2 is as follows: S21: Input any aerial remote sensing target detection image F1 into the DHFaster module, and output the aerial remote sensing target detection feature map F2 by efficiently capturing refined spatial features and contextual information; S22: Then F2 is input into the DHFaster module to further capture refined spatial features and contextual information, and outputs the airborne remote sensing target detection feature map F3; S23: Then input F3 into the MCM module to enhance the feature extraction capability of the target and output the aerial remote sensing target detection feature map F4; S24: Input F4 into the DHFaster module and output the aerial remote sensing target detection feature map F5. Then input F5 into the MCM module and output the aerial remote sensing enhancement feature map P1. S25: Input P1 into the DHFaster module and output the aerial remote sensing target detection feature map F6. Then input F6 into the MCM module and output the aerial remote sensing enhancement feature map P2. S26: Input P2 into the DHFaster module and output the aerial remote sensing target detection feature map F7. Then input F7 into the MCM module and output the aerial remote sensing target detection feature map F8. S27: Finally, input F8 into the SPPF module to generate the aerial remote sensing enhanced feature map P3; S28: After processing by the multi-scale contextual feature enhancement extractor described above, aerial remote sensing enhanced feature maps P1, P2 and P3 are obtained.

2. The target detection method for aerial remote sensing images based on the HSNet model according to claim 1, characterized in that, The specific steps of S1 are as follows: S11: Based on existing public datasets, collect aerial remote sensing images and construct an initial target detection dataset; S12: Preprocess the initial target detection dataset to obtain the preprocessed target detection dataset; S13: Divide the preprocessed target detection dataset to obtain the aerial remote sensing image target detection dataset.

3. The target detection method for aerial remote sensing images based on the HSNet model according to claim 1, characterized in that, The construction and execution process of the DHFaster module in S21 is as follows: S211: Treat the input feature map of the DHFaster module as a context-interactive fusion feature map H1, and process H1 through two branches: In the first branch, H1 is first subjected to a Conv1×1 convolution operation to initially adjust the number of channels and retain the original spatial location information, resulting in a context-interactive fusion feature map H2; then, H2 is subjected to a Conv3×3 convolution operation to extract information on local image features, resulting in a context-interactive fusion feature map H3; finally, H3 is subjected to a Conv1×1 convolution operation to fuse features from different channels at the same location, resulting in a context-interactive fusion feature map H4. In the second branch, H1 is first divided into three groups—Group1, Group2, and Group3—along the channel dimension, allowing different groups to focus on different types of feature patterns. Then, the feature map of each group is further divided into k×k non-overlapping local sub-regions. Each local sub-region uses an independent, non-shared convolutional kernel for feature extraction, ensuring the model can capture and preserve subtle feature differences at each unique location. Afterward, all processed, size-invariant sub-regions are reassembled in their original positions, restoring the feature map to its original size before sub-region division. After these operations, Group1... Group 1 yields context interaction fusion feature map H5_1, Group 2 yields context interaction fusion feature map H5_2, and Group 3 yields context interaction fusion feature map H5_3. Then, H5_1, H5_2, and H5_3 are concatenated along the channel dimension to generate context interaction fusion feature map H5. Next, H5 is convolved with a Conv1×1 convolution to achieve information exchange and fusion among the three groups, resulting in context interaction fusion feature map H6. Finally, H6 is convolved with a Conv3×3 convolution to generate context interaction fusion feature map H7. S212: Finally, perform a Concat operation on H4 and H7 to complete the splicing along the channel dimension and generate a contextual interactive fusion feature map H8; H8 is the output feature map of the DHFaster module.

4. The target detection method for aerial remote sensing images based on the HSNet model according to claim 1, characterized in that, The construction and execution process of the MCM module in S23 is as follows: S231: Treat the input feature map of the MCM module as a multi-domain enhanced fusion feature map X1, and process X1 through three branches: In the first branch, X1 is first subjected to a Conv1×1 convolution operation to obtain a multi-domain enhanced fusion feature map X2; then X2 is subjected to a Conv3×3 convolution operation to extract local spatial features to obtain a multi-domain enhanced fusion feature map X3; then X3 is input into a dilated convolution module DilatedConv with a kernel size of 3×3 and a dilation rate of 1 to further enhance and consolidate the local feature expression to obtain a multi-domain enhanced fusion feature map X4; X1 and X4 are subjected to a Concat operation to complete the concatenation along the channel dimension to enrich the hierarchy of features and generate a multi-domain enhanced fusion feature map X5; In the second branch, X1 is first subjected to a Conv1×1 convolution operation to obtain a multi-domain enhanced fusion feature map X6; then X6 is subjected to a Conv3×3 convolution operation to extract local spatial features, resulting in a multi-domain enhanced fusion feature map X7; X7 is then input into a dilated convolution module DilatedConv with a kernel size of 3×3 and a dilation rate of 3 to capture mid-range contextual information, resulting in a multi-domain enhanced fusion feature map X8; X1 and X8 are then subjected to a Concat operation to complete the concatenation along the channel dimension to enrich the hierarchical nature of the features, generating a multi-domain enhanced fusion feature map X9; In the third branch, X1 is first subjected to a Conv1×1 convolution operation to obtain a multi-domain enhanced fusion feature map X10; then, X10 is subjected to a Conv3×3 convolution operation to extract local spatial features, resulting in a multi-domain enhanced fusion feature map X11; then, X11 is input into a DilatedConv convolution module with a kernel size of 3×3 and a dilation rate of 5 to capture long-range dependencies and global scene context, resulting in a multi-domain enhanced fusion feature map X12; finally, X1 and X12 are subjected to a Concat operation to complete the concatenation along the channel dimension to enrich the hierarchy of features, generating a multi-domain enhanced fusion feature map X13; S232: Subsequently, through three learnable fusion weights Adaptive fusion of X5, X9 and X13 is performed to generate multi-domain enhanced fusion feature map X14; S233: Then, perform a Conv1×1 convolution operation on X14 to adjust the number of channels of X14 to match the number of channels of X1, and obtain the multi-domain enhanced fusion feature map X15; finally, add X1 and X15 element by element to prevent the gradient from vanishing in the deep network, and obtain the multi-domain enhanced fusion feature map X16; X16 is the output feature map of the MCM module.

5. The target detection method for aerial remote sensing images based on the HSNet model according to claim 1, characterized in that, The specific construction process of the adaptive feature refinement fusion unit in S3 is as follows: S31: First, input the enhanced feature map P3 of the airborne remote sensing obtained by the multi-scale context feature enhancement extractor into the DCF_L module to enhance the representation ability of large targets and overall structure, and output the airborne remote sensing target detection feature map F9. S32: Then, upsample F9 to generate the aerial remote sensing target detection feature map F10; then, add the aerial remote sensing enhanced feature map P2 obtained by the multi-scale context feature enhancement extractor and F10 element by element to obtain the aerial remote sensing target detection feature map F11; input F11 into the C2f module to obtain the aerial remote sensing target detection feature map F12; input F12 into the DCF_S module to further enhance the representation ability of small targets in the features, and output the aerial remote sensing target detection feature map F13; then, upsample F13 to generate the aerial remote sensing target detection feature map F14; S33: Next, the aerial remote sensing enhanced feature maps P1 and F14 obtained by the multi-scale context feature enhancement extractor are added element by element to obtain the aerial remote sensing target detection feature map F15; F15 is input into the C2f module to refine the target features and obtain the aerial remote sensing refined fusion feature map R1. S34: Input R1 into the DHFaster module to further capture refined spatial features and contextual information, and output the aerial remote sensing target detection feature map F16; then add F13 and F16 element by element to obtain the aerial remote sensing target detection feature map F17. S35: Input F17 into the C2f module to further refine the target features and obtain the refined fusion feature map R2 of airborne remote sensing; input R2 into the DHFaster module to deeply capture refined spatial features and contextual information and output the airborne remote sensing target detection feature map F18. S36: Then add F9 and F18 element by element to obtain the aerial remote sensing target detection feature map F19; then input F19 into the C2f module to perform deep refinement of the target features and obtain the aerial remote sensing refined fusion feature map R3. S37: After processing by the above adaptive feature refinement fusion processor, the refined and fused feature maps R1, R2 and R3 of airborne remote sensing are obtained.

6. The target detection method for aerial remote sensing images based on the HSNet model according to claim 5, characterized in that, The construction and execution process of the DCF_L module in S31 is as follows: S311: Treat the input feature map of the DCF_L module as a large target cascaded fusion feature map Q1, input Q1 into the multi-branch redundant feature generation module for the large target feature map, and output the large target cascaded fusion feature map Q2. S312: Q2 is then processed through two branches to further process and fuse the target features: In the first branch, MaxPool is first performed on Q2, which uses a 3×3 pooling window that slides with a stride of 1 pixel and fills the edges of Q2 with a ring of pixels to highlight salient features, resulting in the large target cascaded fusion feature map Q3. At the same time, AvgPool is performed on Q2, which uses a 3×3 pooling window that slides with a stride of 1 pixel and fills the edges of Q2 with a ring of pixels to suppress noise and smooth features, resulting in the large target cascaded fusion feature map Q4. Then, Q3 and Q4 are concatted together along the channel dimension to obtain the large target cascaded fusion feature map Q5. Finally, Conv1×1 convolution is performed on Q5 to make the features more compact, resulting in the large target cascaded fusion feature map Q6. Q6 is then input into the Channel Attention Module, which learns to obtain the importance weight vector W1 for different channels, which is used to evaluate the importance of each feature channel. Then, a weighted operation is performed on the channel dimension of Q6 based on W1, that is, W1 and Q6 are multiplied one channel at a time to obtain the large target cascaded fusion feature map Q7. In the second branch, Q2 is first subjected to a Conv3×3 convolution operation to further extract and transform the features, resulting in a large target cascaded fusion feature map Q8. Then, Q8 is subjected to a Conv1×1 convolution operation to obtain a large target cascaded fusion feature map Q9. Next, Q9 is input into the Spatial Attention Module to learn the importance weight matrix W2 of the feature map in the spatial dimension, which is used to evaluate the importance of each spatial position. Then, the spatial positions of Q9 are weighted according to W2, that is, W2 is multiplied with Q9 position by position to obtain a large target cascaded fusion feature map Q10. S313: Finally, perform a Concat operation on Q7 and Q10 to complete the splicing along the channel dimension and obtain the large target cascaded fusion feature map Q11. Perform a Conv1×1 convolution operation on Q11 to obtain the large target cascaded fusion feature map Q12; Q12 is the output feature map of the DCF_L module.

7. The target detection method for aerial remote sensing images based on the HSNet model according to claim 5, characterized in that, The construction and execution process of the DCF_S module in S32 is as follows: S321: Treat the input feature map of the DCF_S module as a small target cascaded fusion feature map Z1, input Z1 into the multi-branch redundant feature generation module for small target feature maps, and output the small target cascaded fusion feature map Z2. S322: Z2 is then processed through two branches to further process and fuse target features: In the first branch, MaxPool is first performed on Z2, which uses a 3×3 pooling window that slides with a stride of 1 pixel and fills the edges of Z2 with a full ring of pixels to highlight salient features, resulting in the small target cascaded fusion feature map Z3. Simultaneously, AvgPool is performed on Z2, which uses a 3×3 pooling window that slides with a stride of 1 pixel and fills the edges of Z2 with a full ring of pixels to suppress noise and smooth features, resulting in the small target cascaded fusion feature map Z4. Then, Z3 and Z4 are concatenated along the channel dimension to obtain the small target cascaded fusion feature map Z5. Finally, Z5 is convolved with a Conv1×1 convolution to make the features more compact, resulting in the small target cascaded fusion feature map Z6. Z6 is then input into the Channel Attention Module to learn the importance weight vector W3 for different channels, which is used to evaluate the importance of each feature channel. Then, a weighted operation is performed on the channel dimension of Z6 based on W3, that is, W3 and Z6 are multiplied one channel at a time to obtain the small target cascaded fusion feature map Z7. In the second branch, Z2 is first subjected to a Conv3×3 convolution operation to further extract and transform the features, resulting in a small target cascaded fusion feature map Z8. Then, Z8 is subjected to a Conv1×1 convolution operation to obtain a small target cascaded fusion feature map Z9. Z9 is then input into the Spatial Attention Module to learn the importance weight matrix W4 of the feature map in the spatial dimension, which is used to evaluate the importance of each spatial location. Then, the spatial locations of Z9 are weighted according to W4, that is, W4 is multiplied with Z9 position by position to obtain a small target cascaded fusion feature map Z10. S323: Finally, perform a Concat operation on Z7 and Z10 to complete the splicing along the channel dimension, and obtain the small target cascaded fusion feature map Z11; perform a Conv1×1 convolution operation on Z11 to obtain the small target cascaded fusion feature map Z12; Z12 is the output feature map of the DCF_S module.

8. The target detection method for aerial remote sensing images based on the HSNet model according to claim 1, characterized in that, The specific steps of S4 are as follows: S41: Input any aerial remote sensing target detection image F1 into the multi-scale context feature enhancement extractor in the HSNet model. Through multi-scale context expansion and spatial detail feature enhancement, aerial remote sensing enhanced feature maps P1, P2 and P3 are obtained. S42: Input P1, P2 and P3 into the adaptive feature refinement fusion unit in the HSNet model to achieve adaptive fusion and refinement of multi-scale target features, and obtain the refined fusion feature maps R1, R2 and R3 of airborne remote sensing. S43: Input R1, R2, and R3 into three independent target detection heads respectively, and each head obtains a prediction tensor containing target detection information from the aerial remote sensing image; Subsequently, the prediction results of the three heads are merged together, and the high-confidence detection boxes are filtered out by the non-maximum suppression algorithm (NMS). Finally, the detection result image with target annotation is output.

Images