Feature extraction and dynamic fusion method and system for remote sensing small target detection

Abstract

The invention belongs to the technical field of remote sensing image processing, and particularly relates to a remote sensing small target detection-oriented feature extraction and dynamic fusion method and system, and the method comprises the steps: constructing a remote sensing small target detection model, building a wavelet multi-scale feature extraction module in a model backbone network, and enabling the wavelet multi-scale feature extraction module to comprise a global branch and a local branch, the global branch is used for introducing self-attention, a channel and a space attention mechanism to extract global context and structure prior and enhancing horizontal and vertical detail components to improve edge expression, the local branch is used for extracting multi-scale local features, and output features of the global branch and the local branch are spliced in a channel dimension to obtain a multi-scale local feature; in this way, a hierarchical feature representation covering global-local-fine granularity is formed. According to the remote sensing small target detection model, key information in a remote sensing image can be accurately extracted, and confusion between the small target and the background is reduced, so that the detection precision of the remote sensing small target is effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing technology, specifically relating to a feature extraction and dynamic fusion method and system for remote sensing small target detection. Background Technology

[0002] Remote sensing target detection based on satellites and UAVs has become a crucial support for modern traffic monitoring. However, traffic targets in remote sensing images are typically small in size and easily affected by complex backgrounds, posing significant challenges to detection accuracy and robustness. Therefore, conducting research on the accurate detection of small remote sensing targets is of great importance for improving the intelligence and safety of traffic supervision.

[0003] In recent years, the development of deep learning, especially convolutional neural networks (CNNs), has greatly promoted the progress of small object detection. Compared with traditional methods, deep learning has shown significant advantages in robustness and adaptability. In remote sensing small object detection tasks, remote sensing images not only contain rich details and textures but also important global structural information. Deep learning models extract and fuse multi-scale features to capture local details and overall layout in images, thereby achieving a deeper understanding of image content and improving the accuracy of small object recognition. Some studies have enhanced the model's ability to extract multi-scale features by improving the receptive field of CNNs. To further improve the extraction of global spatial structure of images by deep learning models during multi-scale feature modeling, enabling the model to more accurately grasp the overall spatial layout of images, some studies have introduced attention mechanisms into deep learning models. In addition, multi-scale feature fusion is crucial for the model to understand image content and accurately identify target objects. Most studies use feature pyramid networks (FPNs) for multi-scale feature fusion, and some studies have improved feature fusion techniques based on this, further enhancing the integration effect of multi-scale features.

[0004] In remote sensing scenarios, small targets are typically small in size, occupying only a few pixels, and often exist in environments with low contrast and strong background interference, increasing the difficulty of target detection. Although existing deep learning-based methods improve the model's understanding of images by introducing multi-scale feature processing, thereby improving the detection accuracy of small targets, there are certain limitations in multi-scale feature extraction and fusion. Especially when the contrast between the target and the background is low, the following two challenges still exist, making it difficult to accurately identify small targets:

[0005] First, current methods typically extract relatively simple and highly varied features when extracting multi-scale features, which makes it difficult for the model to fully understand the relationship between the target and the background, thus causing confusion between the target and the background.

[0006] Secondly, since small targets account for a small proportion of information in the whole image, existing methods have failed to effectively integrate their key information during multi-scale feature fusion, resulting in the dilution of important features in the network neck and downstream fusion stages, which further limits the separability and detectability of small targets. Summary of the Invention

[0007] The purpose of this invention is to address the aforementioned problems in the existing technology by providing a feature extraction and dynamic fusion method and system for remote sensing small target detection that can extract hierarchical feature representations covering global, local, and fine-grained levels, thereby reducing confusion between small targets and the background and effectively improving the detection accuracy of remote sensing small targets.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows:

[0009] In a first aspect, the present invention provides a feature extraction and dynamic fusion method for remote sensing small target detection, the feature extraction and dynamic fusion method comprising:

[0010] S1. Construct a remote sensing small target detection model; the remote sensing small target detection model includes a backbone network, a neck network and a detection head. The backbone network has a built-in wavelet multi-scale feature extraction module for multi-scale feature extraction. The neck network has a built-in dynamic feature fusion module for fusing the extracted multi-scale features. The detection head is used for target detection based on the fused multi-scale features.

[0011] The wavelet multi-scale feature extraction module includes a global branch and a local branch. The global branch decomposes the input feature map into four different frequency domain components using Haar wavelet transform: a low-frequency approximation component, a horizontal component, a vertical component, and a diagonal component. For the low-frequency approximation component, a self-attention mechanism is used to obtain the query, key, and value. Then, spatial attention and channel attention weighting are applied to both the query and key to obtain the query-weighted result and the key-weighted result. Simultaneously, the diagonal component is discarded, and the horizontal and vertical components are added to the value. Finally, the value summation result is fused with the query-weighted result and the key-weighted result to obtain the output feature of the global branch. The local branch extracts multi-scale local features from the input feature map. The output features of the global branch and the local branch are concatenated along the channel dimension to generate the final output feature of the wavelet multi-scale feature extraction module.

[0012] S2. Collect high-resolution remote sensing images to construct a training dataset. Train the remote sensing small target detection model based on the training dataset, and use the trained remote sensing small target detection model to achieve target detection.

[0013] The mathematical expression for the global branch is:

[0014] ;

[0015] ;

[0016] ;

[0017] ;

[0018] ;

[0019] ;

[0020] In the above formula, , These represent the input and output features of the global branch, respectively. , , , , , Representing Haar wavelet transform, channel segmentation, and Convolutional operations, spatial attention mechanisms, channel attention mechanisms, and upsampling operations; , , , These represent the low-frequency approximate components, horizontal components, vertical components, and diagonal components obtained through Haar wavelet transform, respectively. , , These are the query, key, and value obtained after introducing a self-attention mechanism for the low-frequency approximate component, respectively. , , These are the query weighted results, key weighted results, and value sum results, respectively.

[0021] The local branch is a multi-scale depthwise separable convolutional parallel architecture, which includes three parallel branches. The three branches are depthwise separable convolutional branches with 3×3 convolutional kernels, 5×5 convolutional kernels, and 7×7 convolutional kernels, respectively.

[0022] The mathematical expression for the local branch is:

[0023] ;

[0024] In the above formula, , These represent the input and output features of the local branch, respectively. , , These represent depthwise separable convolutions with 3×3, 5×5, and 7×7 kernels, respectively.

[0025] The dynamic feature fusion module includes point convolution, convolution RFAConv, spatial shift, and adaptive weighting mechanisms. First, point convolution expands the channel dimension of the input features to the dynamic feature fusion module, resulting in an expanded feature map. Then, the expanded feature map is segmented into features according to the channel dimension. , , , for features Feature extraction is performed using convolutional RFAConv, and the extracted features are then compared with the features. The first intermediate feature is obtained by weighted fusion using an adaptive weighting mechanism. , the first intermediate feature ,feature Then, feature extraction is performed using RFAConv convolution, and the extracted feature results are concatenated by channels to obtain the target's external contour features. ; target external contour features With features The second intermediate feature is obtained by weighted fusion using an adaptive weighting mechanism. ; for the second intermediate feature First, positional information is injected through a spatial displacement module, and then point convolution is used to process the second intermediate features. Perform fine-grained texture feature extraction; combine the extracted texture features with the relevant data. The internal texture features of the target are obtained by weighted fusion using an adaptive weighting mechanism. Finally, the target's external contour features are... With target internal texture features Channel splicing is performed to obtain the output features of the dynamic feature fusion module.

[0026] The mathematical expression for the dynamic feature fusion module is:

[0027] ;

[0028] ;

[0029] ;

[0030] ;

[0031] ;

[0032] ;

[0033] In the above formula, , These represent the input and output features of the dynamic feature fusion module, respectively. , , , These represent channel segmentation, pointwise convolution, channel concatenation, and convolution RFAConv, respectively. For spatial displacement operations; , , , , , These are the dynamic weights of the corresponding features; , These are the first intermediate feature and the second intermediate feature, respectively. , These are the target's external contour features and internal texture features extracted by the dynamic feature fusion module, respectively.

[0034] Secondly, the present invention provides a feature extraction and dynamic fusion system for remote sensing small target detection, the feature extraction and dynamic fusion system comprising:

[0035] The model building module is used to build a remote sensing small target detection model. The remote sensing small target detection model includes a backbone network, a neck network, and a detection head. The backbone network has a built-in wavelet multi-scale feature extraction module for multi-scale feature extraction. The neck network has a built-in dynamic feature fusion module for fusing the extracted multi-scale features. The detection head is used for target detection based on the fused multi-scale features.

[0036] The wavelet multi-scale feature extraction module includes a global branch and a local branch. The module concatenates the output features of the global and local branches along the channel dimension to generate the final output features. The global branch decomposes the input feature map into four different frequency domain components using Haar wavelet transform: a low-frequency approximation component, a horizontal component, a vertical component, and a diagonal component. For the low-frequency approximation component, a self-attention mechanism is used to obtain the query, key, and value. Then, spatial attention and channel attention weighting are applied to both the query and key to obtain the query-weighted result and the key-weighted result. Simultaneously, the diagonal component is discarded, and the horizontal and vertical components are added to the value. Finally, the value summation result is fused with the query-weighted result and the key-weighted result to obtain the output features of the global branch. The local branch extracts multi-scale local features from the input feature map.

[0037] The model training and target detection module is used to collect high-resolution remote sensing images to build a training dataset, train the remote sensing small target detection model based on the training dataset, and use the trained remote sensing small target detection model to achieve target detection.

[0038] The mathematical expression for the global branch is:

[0039] ;

[0040] ;

[0041] ;

[0042] ;

[0043] ;

[0044] ;

[0045] In the above formula, , These represent the input and output features of the global branch, respectively. , , , , , Representing Haar wavelet transform, channel segmentation, and Convolutional operations, spatial attention mechanisms, channel attention mechanisms, and upsampling operations; , , , These represent the low-frequency approximate components, horizontal components, vertical components, and diagonal components obtained through Haar wavelet transform, respectively. , , These are the query, key, and value obtained after introducing a self-attention mechanism for the low-frequency approximate component, respectively. , , These are the query weighted results, key weighted results, and value sum results, respectively.

[0046] The local branch is a multi-scale depthwise separable convolutional parallel architecture, which includes three parallel branches. The three branches are depthwise separable convolutional branches with 3×3 convolutional kernels, 5×5 convolutional kernels, and 7×7 convolutional kernels, respectively.

[0047] The mathematical expression for the local branch is:

[0048] ;

[0049] In the above formula, , These represent the input and output features of the local branch, respectively. , , These represent depthwise separable convolutions with 3×3, 5×5, and 7×7 kernels, respectively.

[0050] The dynamic feature fusion module includes point convolution, RFAConv convolution, spatial shift, and an adaptive weighting mechanism. The dynamic feature fusion module performs dynamic feature fusion according to the following steps: first, it expands the channel dimension of the input features by point convolution to obtain an expanded feature map; then, it segments the expanded feature map into features according to the channel dimension. , , , for features Feature extraction is performed using convolutional RFAConv, and the extracted features are then compared with the features. The first intermediate feature is obtained by weighted fusion using an adaptive weighting mechanism. , the first intermediate feature ,feature Then, feature extraction is performed using RFAConv convolution, and the extracted feature results are concatenated by channels to obtain the target's external contour features. ; target external contour features With features The second intermediate feature is obtained by weighted fusion using an adaptive weighting mechanism. ; for the second intermediate feature First, positional information is injected through a spatial displacement module, and then point convolution is used to process the second intermediate features. Perform fine-grained texture feature extraction; combine the extracted texture features with the relevant data. The internal texture features of the target are obtained by weighted fusion using an adaptive weighting mechanism. Finally, the target's external contour features are... With target internal texture features Channel splicing is performed to obtain the output features of the dynamic feature fusion module.

[0051] The mathematical expression for the dynamic feature fusion module is:

[0052] ;

[0053] ;

[0054] ;

[0055] ;

[0056] ;

[0057] ;

[0058] In the above formula, , These represent the input and output features of the dynamic feature fusion module, respectively. , , , These represent channel segmentation, pointwise convolution, channel concatenation, and convolution RFAConv, respectively. For spatial displacement operations; , , , , , These are the dynamic weights of the corresponding features; , These are the first intermediate feature and the second intermediate feature, respectively. , These are the target's external contour features and internal texture features extracted by the dynamic feature fusion module, respectively.

[0059] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0060] 1. In the feature extraction and dynamic fusion method for remote sensing small target detection described in this invention, a wavelet multi-scale feature extraction module is built into the backbone network. Multi-scale feature extraction is performed by combining self-attention mechanism and multi-branch convolutional network. Specifically, the wavelet multi-scale feature extraction module includes a global branch and a local branch. The global branch is used to introduce self-attention and channel and spatial attention mechanisms to extract global context and structural priors, and to enhance horizontal and vertical detail components to improve edge expression. The local branch is used to extract multi-scale local features through multi-branch convolution. Different branches use convolution kernels with different proportions of receptive field to collaboratively capture pixel-level details, local structure and neighboring background, and suppress the feature oversmoothing problem caused by single-scale convolution. The output features of the global branch and the local branch are concatenated and normalized in the channel dimension to form a hierarchical feature representation covering global-local-fine-grained levels. This is used as the final output feature of the wavelet multi-scale feature extraction module, which can accurately extract key information in remote sensing images and reduce the confusion between small targets and background, thereby effectively improving the detection accuracy of remote sensing small targets.

[0061] 2. In the feature extraction and dynamic fusion method for remote sensing small target detection described in this invention, a dynamic feature fusion module is built into the neck network. The dynamic feature fusion module includes point convolution, convolution RFAConv, spatial displacement, and an adaptive weighting mechanism. Point convolution is used to expand the channel dimension of the input feature map, convolution RFAConv is used to extract coarse-grained contour information of small targets, and the adaptive weighting mechanism is used to dynamically adjust the fusion ratio of coarse-grained and fine-grained information according to the contribution of different features, so as to prevent the loss of target information and feature dilution, and ultimately improve the detection capability of remote sensing small targets in complex backgrounds, especially the preservation of edge and texture information of small targets. Attached Figure Description

[0062] Figure 1 This is a flowchart of the feature extraction and dynamic fusion method for remote sensing small target detection described in this invention.

[0063] Figure 2 This is a network structure diagram of the remote sensing small target detection model constructed in this invention.

[0064] Figure 3 This is a schematic diagram of the wavelet multi-scale feature extraction module in the remote sensing small target detection model constructed in this invention.

[0065] Figure 4 This is a schematic diagram of the dynamic feature fusion module in the remote sensing small target detection model constructed in this invention.

[0066] Figure 5 This is a partial detection result of the remote sensing small target detection model constructed in this invention on the NWPU VHR1.0 dataset.

[0067] Figure 6 This is a partial detection result of the remote sensing small target detection model constructed in this invention on the USOD dataset.

[0068] Figure 7 This is a structural block diagram of the feature extraction and dynamic fusion system for remote sensing small target detection described in this invention. Detailed Implementation

[0069] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings.

[0070] Example 1:

[0071] See Figure 1 A feature extraction and dynamic fusion method for small target detection in remote sensing is proposed, which proceeds in the following steps:

[0072] S1. Construct a remote sensing small target detection model; such as... Figure 2As shown, the remote sensing small target detection model includes a backbone network, a neck network, and a detection head. The remote sensing image to be detected is first processed by the backbone network. The backbone network has a built-in wavelet multi-scale feature extraction module (WMFEM) for multi-scale feature extraction of the remote sensing image to construct multi-scale features of global scene, local structure, and fine-grained texture, helping the system understand the target and background information in the image. In addition, a triple attention module (TA) is added to the backbone network. This module captures and fuses the interaction relationship between the three dimensions of channel, height, and width through three parallel branches, thereby improving the network performance without increasing too many parameters. The network focuses more on more useful channels and spatial locations; then, the extracted features are passed to the neck network, which has a built-in dynamic feature fusion module (DFFM) for adaptive fusion according to the differentiated needs of the extracted multi-scale features. That is, according to the differences in the external contour and internal texture of the target, the dynamic weight ratio of different scale features is flexibly adjusted, so as to better preserve the detailed information of small targets and help reduce the confusion between targets and backgrounds when dealing with small targets and complex backgrounds. The detection head is used to perform target detection based on the fused multi-scale features and output the target type and location in the remote sensing image, thereby achieving accurate target recognition.

[0073] Specifically, such as Figure 3 As shown, the wavelet multi-scale feature extraction module includes a global branch and a local branch. The output features of the global branch and the local branch are concatenated along the channel dimension to generate the final output features of the wavelet multi-scale feature extraction module. The global branch is used to decompose the input feature map into four different frequency domain components in some channels of the input feature map using Haar wavelet transform: low-frequency approximate component, horizontal component, vertical component, and diagonal component. Then, for the low-frequency approximate component, the query, key, and value are obtained through a self-attention mechanism. Then, spatial attention and channel attention are applied to both the query and the key to obtain the query weighted result and the key weighted result. At the same time, the diagonal component is discarded, and the horizontal and vertical components are added to the value. Finally, the value summation result is fused with the query weighted result and the key weighted result to obtain the output features of the global branch.

[0074] The low-frequency approximation component contains the main structural information of the image, while the horizontal, vertical, and diagonal components capture the edge information of the image, respectively. A self-attention mechanism is introduced into the low-frequency approximation component. This design overcomes the limitations of the local receptive field in traditional convolutional operations, enabling the extraction of richer global contextual information from the image. By applying spatial and channel attention weights to the query and key, the model can highlight key regions and channels, thereby enhancing the extraction of global contextual information. The horizontal and vertical components are added to the value, and the diagonal component is discarded to further enhance the representation of detailed information and reduce noise interference.

[0075] For the input feature map The mathematical expression for the global branch is:

[0076] ;

[0077] ;

[0078] ;

[0079] ;

[0080] ;

[0081] ;

[0082] In the above formula, , These represent the input and output features of the global branch, respectively. , , , , , Representing Haar wavelet transform, channel segmentation, and Convolutional operations, spatial attention mechanisms, channel attention mechanisms, and upsampling operations; , , , These represent the low-frequency approximate components, horizontal components, vertical components, and diagonal components obtained through Haar wavelet transform, respectively. , , These are the query, key, and value obtained after introducing a self-attention mechanism for the low-frequency approximate component, respectively. , , These are the query weighted results, key weighted results, and value sum results, respectively.

[0083] Specifically, the local branch is a multi-scale depthwise separable convolutional parallel architecture, including three parallel branches with different receptive fields. These three branches are depthwise separable convolutional branches with 3×3, 5×5, and 7×7 convolutional kernels, respectively. While maintaining the high efficiency of depthwise separable convolution parameters, it achieves the collaborative capture of multi-granular local texture information on the remaining channels of the input feature map. The 3×3 kernel depthwise separable convolutional branch focuses on extracting pixel-level fine-grained features and can effectively capture texture details; the 5×5 kernel depthwise separable convolutional branch expands the receptive field to a medium range and can establish medium-range spatial correlations between local features; while the 7×7 kernel depthwise separable convolutional branch can perceive a larger range of local background information, providing background support for feature understanding. Through the three parallel branches, multi-level feature representation from pixel-level details to local background is achieved, which can effectively solve the problem of feature oversmoothing caused by single-scale convolutional kernels.

[0084] For the input feature map The mathematical expression for the local branch is:

[0085] ;

[0086] In the above formula, , These represent the input and output features of the local branch, respectively. , , These represent depthwise separable convolutions with 3×3, 5×5, and 7×7 kernels, respectively.

[0087] Specifically, such as Figure 4 As shown, the dynamic feature fusion module includes point convolution, RFAConv convolution, spatial shifting, and an adaptive weighting mechanism. Specifically, it first expands the channel dimension of the input features of the dynamic feature fusion module through point convolution to provide sufficient feature space for subsequent processing; then, the expanded feature map is segmented into features according to the channel dimension. , , , for features Feature extraction is performed using convolutional RFAConv, and the extracted features are then compared with the features. The first intermediate feature is obtained by weighted fusion using an adaptive weighting mechanism. , the first intermediate feature ,feature The input is then fed into RFAConv for feature extraction, and the extracted features are concatenated by channels to obtain the target's external contour features. ; target external contour features With features The second intermediate feature is obtained by weighted fusion using an adaptive weighting mechanism. ; for the second intermediate feature First, positional information is injected through a spatial displacement module, thus providing spatial information extraction capabilities for subsequent point convolution; then, point convolution is used to process the second intermediate features. Fine-grained texture feature extraction is performed to effectively address the problem of detail dilution caused by excessively large receptive fields in convolutional neural networks; the texture feature extraction results and features are then analyzed. The internal texture features of the target are obtained by weighted fusion using an adaptive weighting mechanism. Finally, the target's external contour features are... With target internal texture features Channel splicing is performed to obtain the output features of the dynamic feature fusion module.

[0088] By leveraging external contour features and internal texture features of the target, the contour blurring problem caused by parameter sharing in convolutional neural networks is effectively overcome. An adaptive weighting mechanism is adopted during the channel stitching process. Based on the importance of the external contour information and internal texture features of the small target, the fusion weight of coarse-grained contour and fine-grained texture information is dynamically adjusted, which can effectively preserve the detailed information of the small target, thereby preventing information loss and feature dilution.

[0089] for The mathematical expression for the dynamic feature fusion module is:

[0090] ; ;

[0091] ;

[0092] ; ;

[0093] ;

[0094] In the above formula, , These represent the input and output features of the dynamic feature fusion module, respectively. , , , These represent channel segmentation, pointwise convolution, channel concatenation, and convolution RFAConv, respectively. For spatial displacement operations; , , , , , These are the dynamic weights of the corresponding features; , These are the first intermediate feature and the second intermediate feature, respectively. , These are the target's external contour features and internal texture features extracted by the dynamic feature fusion module, respectively.

[0095] S2. Collect high-resolution remote sensing images to construct a training dataset. Train the remote sensing small target detection model based on the training dataset, and use the trained remote sensing small target detection model to achieve target detection.

[0096] Performance verification:

[0097] 1. To verify the superiority of the proposed method, performance comparison experiments were conducted on the following two publicly available test datasets: (1) The NWPU VHR1.0 dataset is designed for target detection tasks in high-resolution remote sensing images. This dataset contains 800 images covering 10 common categories: aircraft (A), ships (S), oil tanks (ST), baseball fields (BF), badminton courts (TC), basketball courts (BC), runways (GTR), harbors (H), bridges (B), and vehicles (V). (2) The USOD dataset is derived from the UNICORN 2008 dataset and is extracted by selecting samples with small target features in remote sensing applications. This dataset contains 3,000 images and 43,378 labeled vehicle instances, of which 96.3% of the objects are smaller than 16×16 pixels and 99.9% are smaller than 32×32 pixels. This dataset includes many instances captured under challenging conditions, such as low light and shadow occlusion. In addition, various image degradation factors were considered, including blur, additive Gaussian noise, stripe noise, and haze.

[0098] Implementation details and evaluation metrics: (1) Implementation details: The method proposed in this invention is implemented based on the PyTorch framework, using the SGD optimizer, with an initial learning rate of 0.01 and a momentum of 0.937. During training, all test datasets are divided into training and validation sets in an 8:2 ratio. The experiment is conducted on an RTX 4090D GPU, with a batch size of 8 and a training cycle of 250 rounds. (2) Evaluation metrics: Average precision (mAP) is used as the main evaluation metric. mAP is calculated by averaging the AP values ​​of all classes to obtain the overall performance of the model on various targets.

[0099] The comparison results of the proposed method with other models on the NWPU VHR1.0 dataset are shown in Table 1:

[0100] Table 1. Comparison results of the proposed method with other models on the NWPU VHR1.0 dataset.

[0101]

[0102] As shown in Table 1, the method proposed in this invention has 6.68M parameters, which is smaller than existing models (e.g., VanillaNet's 29.54M and Hyper-YOLO's 30.70M). Simultaneously, it achieves a mean accuracy (mAP) of 93.2% on the NWPU VHR1.0 dataset, an improvement over YOLOv11's 91.9%. In specific target detection performance, the accuracy for oil tank detection is 98.5%, bridge detection is 82.4%, and athletic field detection is 99.4%. In traffic facility detection, the accuracy for port detection is 98.2%, aircraft detection is 99.4%, and vehicle detection is 90.2%. These experimental results demonstrate that the method proposed in this invention exhibits good accuracy in detecting small remote sensing targets such as oil tanks, while achieving high detection performance while maintaining a small parameter count. Figure 5 This paper presents partial detection results of the proposed method on the NWPU VHR1.0 dataset. Figure 5 The detection targets in Figures a and b are an airplane and a harbor, respectively.

[0103] The comparison results of the proposed method with other models on the USOD dataset are shown in Table 2:

[0104] Table 2. Comparison results of the proposed method with other models on the USOD dataset.

[0105]

[0106] As shown in Table 2, the proposed method achieves an mAP of 88.5% on the USOD dataset, demonstrating superior detection performance. Although Hyper-YOLO achieves an mAP of 90.3% with 30.69M parameters, its model complexity is 4.6 times that of the proposed method, which may lead to higher computational resource requirements in practical deployments. The experimental results show that the proposed method effectively improves the feature representation capability for small targets, and through reasonable structural design, it can provide excellent detection performance while ensuring low computational burden. Figure 6 This paper presents some detection results of the proposed method on the USOD dataset. Figure 6 Figures a and b show the detection results for vehicle vehicles in different samples.

[0107] In summary, the method proposed in this invention outperforms existing technologies on test datasets such as NWPU VHR1.0 and USOD, significantly improving the detection accuracy and robustness of small targets, especially showing a significant advantage in the detection of key small targets such as oil tanks and ships.

[0108] 2. To verify the effectiveness of the wavelet multi-scale feature extraction module (WMFEM) and dynamic feature fusion module (DFFM) in the proposed method, ablation experiments were conducted on the NWPU VHR1.0 dataset. The ablation experiment results are shown in Table 3:

[0109] Table 3 Ablation Experiment Results

[0110]

[0111] Table 3 presents the ablation experiment results on the NWPU VHR1.0 dataset. First, the baseline model achieves an mAP of 89.0%. Introducing the Wavelet Multi-Scale Feature Extraction (WMFEM) module alone improves the mAP to 90.7%. Specifically, WMFEM improves detection accuracy in some key categories (such as ST and BC), primarily due to its crucial role in helping the model distinguish targets from the background. Introducing the Dynamic Feature Fusion (DFFM) module alone further enhances the baseline model's mAP to 91.2%. This improvement indicates that DFFM plays a key role in reducing information loss in small targets and helps the model achieve more accurate detection results in certain categories (such as ST). Finally, the complete model combining WMFEM and DFFM (Baseline + WMFEM + DFFM) achieves a maximum mAP of 93.2%, demonstrating the synergistic effect of these two modules in improving the overall model performance. Specifically, the complete model further improves detection accuracy in multiple categories (such as athletic fields, oil tanks, and aircraft), particularly excelling in the challenging task of small target detection.

[0112] Example 2:

[0113] See Figure 7A feature extraction and dynamic fusion system for remote sensing small target detection is disclosed, comprising a model building module, a model training module, and a target detection module. The model building module is used to construct a remote sensing small target detection model. The remote sensing small target detection model includes a backbone network, a neck network, and a detection head. The backbone network incorporates a wavelet multi-scale feature extraction module for multi-scale feature extraction. The neck network incorporates a dynamic feature fusion module for fusing the extracted multi-scale features. The detection head is used for target detection based on the fused multi-scale features. The wavelet multi-scale feature extraction module includes a global branch and a local branch, which are used to combine the global branch and the local branch. The output features of the global branch are concatenated along the channel dimension to generate the final output features of the wavelet multi-scale feature extraction module. The global branch is used to decompose the input feature map into four different frequency domain components through Haar wavelet transform: low-frequency approximate component, horizontal component, vertical component, and diagonal component. Then, for the low-frequency approximate component, the query, key, and value are obtained through a self-attention mechanism. Then, spatial attention and channel attention are applied to both the query and key to obtain the query weighted result and the key weighted result. At the same time, the diagonal component is discarded, and the horizontal and vertical components are added to the value. Finally, the value sum is fused with the query weighted result and the key weighted result to obtain the output features of the global branch. The mathematical expression of the global branch is:

[0114] ;

[0115] ;

[0116] ;

[0117] ;

[0118] ;

[0119] ;

[0120] In the above formula, , These represent the input and output features of the global branch, respectively. , , , , , Representing Haar wavelet transform, channel segmentation, and Convolutional operations, spatial attention mechanisms, channel attention mechanisms, and upsampling operations; , , , These represent the low-frequency approximate components, horizontal components, vertical components, and diagonal components obtained through Haar wavelet transform, respectively. , , These are the query, key, and value obtained after introducing a self-attention mechanism for the low-frequency approximate component, respectively. , , These are the query weighted results, key-weighted results, and value-sum results, respectively.

[0121] The local branch is a multi-scale depthwise separable convolutional parallel architecture used to extract multi-scale local features. Specifically, the local branch includes three parallel branches, which are depthwise separable convolutional branches with 3×3, 5×5, and 7×7 convolutional kernels, respectively. The mathematical expression of the local branch is:

[0122] ;

[0123] In the above formula, , These represent the input and output features of the local branch, respectively. , , These represent depthwise separable convolutions with 3×3, 5×5, and 7×7 kernels, respectively.

[0124] The dynamic feature fusion module includes point convolution, RFAConv convolution, spatial shift, and an adaptive weighting mechanism. The dynamic feature fusion module performs dynamic feature fusion according to the following steps: first, it expands the channel dimension of the input features by point convolution to obtain an expanded feature map; then, it segments the expanded feature map into features according to the channel dimension. , , , for features Feature extraction is performed using convolutional RFAConv, and the extracted features are then compared with the features. The first intermediate feature is obtained by weighted fusion using an adaptive weighting mechanism. , the first intermediate feature ,feature Then, feature extraction is performed using RFAConv convolution, and the extracted feature results are concatenated by channels to obtain the target's external contour features. ; target external contour features With features The second intermediate feature is obtained by weighted fusion using an adaptive weighting mechanism. ; for the second intermediate feature First, positional information is injected through a spatial displacement module, and then point convolution is used to process the second intermediate features. Perform fine-grained texture feature extraction; combine the extracted texture features with the relevant data. The internal texture features of the target are obtained by weighted fusion using an adaptive weighting mechanism. Finally, the target's external contour features are... With target internal texture features Channel concatenation is performed to obtain the output features of the dynamic feature fusion module; the mathematical expression of the dynamic feature fusion module is:

[0125] ;

[0126] ;

[0127] ;

[0128] ;

[0129] ;

[0130] ;

[0131] In the above formula, , These represent the input and output features of the dynamic feature fusion module, respectively. , , , These represent channel segmentation, pointwise convolution, channel concatenation, and convolution RFAConv, respectively. For spatial displacement operations; , , , , , These are the dynamic weights of the corresponding features; , These are the first intermediate feature and the second intermediate feature, respectively. , These are the target's external contour features and internal texture features extracted by the dynamic feature fusion module, respectively.

[0132] The model training and target detection module is used to collect high-resolution remote sensing images to construct a training dataset, train the remote sensing small target detection model based on the training dataset, and use the trained remote sensing small target detection model to achieve target detection.

[0133] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program goods. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0134] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0135] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0136] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0137] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A feature extraction and dynamic fusion method for remote sensing small target detection, characterized by: The feature extraction and dynamic fusion method includes: S1. Construct a remote sensing small target detection model; the remote sensing small target detection model includes a backbone network, a neck network and a detection head. The backbone network has a built-in wavelet multi-scale feature extraction module for multi-scale feature extraction. The neck network has a built-in dynamic feature fusion module for fusing the extracted multi-scale features. The detection head is used for target detection based on the fused multi-scale features. The wavelet multi-scale feature extraction module includes a global branch and a local branch. The global branch decomposes the input feature map into four different frequency domain components using Haar wavelet transform: a low-frequency approximation component, a horizontal component, a vertical component, and a diagonal component. For the low-frequency approximation component, a self-attention mechanism is used to obtain the query, key, and value. Then, spatial attention and channel attention weighting are applied to both the query and key to obtain the query-weighted result and the key-weighted result. Simultaneously, the diagonal component is discarded, and the horizontal and vertical components are added to the value. Finally, the value summation result is fused with the query-weighted result and the key-weighted result to obtain the output feature of the global branch. The local branch extracts multi-scale local features from the input feature map. The output features of the global branch and the local branch are concatenated along the channel dimension to generate the final output feature of the wavelet multi-scale feature extraction module. S2. Collect high-resolution remote sensing images to construct a training dataset. Train the remote sensing small target detection model based on the training dataset, and use the trained remote sensing small target detection model to achieve target detection.

2. The feature extraction and dynamic fusion method for remote sensing small target detection according to claim 1, characterized in that: The mathematical expression for the global branch is: ； ； ； ； ； ； In the above formula, , These represent the input and output features of the global branch, respectively. , , , , , Representing Haar wavelet transform, channel segmentation, and Convolutional operations, spatial attention mechanisms, channel attention mechanisms, and upsampling operations; , , , These represent the low-frequency approximate components, horizontal components, vertical components, and diagonal components obtained through Haar wavelet transform, respectively. , , These are the query, key, and value obtained after introducing a self-attention mechanism for the low-frequency approximate component, respectively. , , These are the query weighted results, key weighted results, and value sum results, respectively.

3. The feature extraction and dynamic fusion method for remote sensing small target detection according to claim 1, characterized in that: The local branch is a multi-scale depthwise separable convolutional parallel architecture, which includes three parallel branches. The three branches are depthwise separable convolutional branches with 3×3 convolutional kernels, 5×5 convolutional kernels, and 7×7 convolutional kernels, respectively. The mathematical expression for the local branch is: ； In the above formula, , These represent the input and output features of the local branch, respectively. , , These represent depthwise separable convolutions with 3×3, 5×5, and 7×7 kernels, respectively.

4. The feature extraction and dynamic fusion method for remote sensing small target detection according to claim 1, characterized in that: The dynamic feature fusion module includes point convolution, convolution RFAConv, spatial shift, and adaptive weighting mechanisms. First, point convolution expands the channel dimension of the input features to the dynamic feature fusion module, resulting in an expanded feature map. Then, the expanded feature map is segmented into features according to the channel dimension. , , , for features Feature extraction is performed using convolutional RFAConv, and the extracted features are then compared with the features. The first intermediate feature is obtained by weighted fusion using an adaptive weighting mechanism. , the first intermediate feature ,feature Then, feature extraction is performed using RFAConv convolution, and the extracted feature results are concatenated by channels to obtain the target's external contour features. ; target external contour features With features The second intermediate feature is obtained by weighted fusion using an adaptive weighting mechanism. ; for the second intermediate feature First, positional information is injected through a spatial displacement module, and then point convolution is used to process the second intermediate features. Perform fine-grained texture feature extraction; combine the extracted texture features with the relevant data. The internal texture features of the target are obtained by weighted fusion using an adaptive weighting mechanism. Finally, the target's external contour features are... With target internal texture features Channel splicing is performed to obtain the output features of the dynamic feature fusion module.

5. The feature extraction and dynamic fusion method for remote sensing small target detection according to claim 4, characterized in that: The mathematical expression for the dynamic feature fusion module is: ； ； ； ； ； ； In the above formula, , These represent the input and output features of the dynamic feature fusion module, respectively. , , , These represent channel segmentation, pointwise convolution, channel concatenation, and convolution RFAConv, respectively. For spatial displacement operations; , , , , , These are the dynamic weights of the corresponding features; , These are the first intermediate feature and the second intermediate feature, respectively. , These are the target's external contour features and internal texture features extracted by the dynamic feature fusion module, respectively.

6. A feature extraction and dynamic fusion system for remote sensing small target detection, characterized in that: The feature extraction and dynamic fusion system includes: The model building module is used to build a remote sensing small target detection model. The remote sensing small target detection model includes a backbone network, a neck network, and a detection head. The backbone network has a built-in wavelet multi-scale feature extraction module for multi-scale feature extraction. The neck network has a built-in dynamic feature fusion module for fusing the extracted multi-scale features. The detection head is used for target detection based on the fused multi-scale features. The wavelet multi-scale feature extraction module includes a global branch and a local branch. The module concatenates the output features of the global and local branches along the channel dimension to generate the final output features. The global branch decomposes the input feature map into four different frequency domain components—a low-frequency approximation component, a horizontal component, a vertical component, and a diagonal component—using Haar wavelet transform on some channels. For the low-frequency approximation component, a self-attention mechanism is used to obtain the query, key, and value. Then, spatial attention and channel attention weighting are applied to both the query and key to obtain the query-weighted result and the key-weighted result. Simultaneously, the diagonal component is discarded, and the horizontal and vertical components are added to the value. Finally, the value summation result is fused with the query-weighted result and the key-weighted result to obtain the output features of the global branch. The local branch extracts multi-scale local features from the input feature map. The model training and target detection module is used to collect high-resolution remote sensing images to build a training dataset, train the remote sensing small target detection model based on the training dataset, and use the trained remote sensing small target detection model to achieve target detection.

7. The feature extraction and dynamic fusion system for remote sensing small target detection according to claim 6, characterized in that: The mathematical expression for the global branch is: ； ； ； ； ； ； In the above formula, , These represent the input and output features of the global branch, respectively. , , , , , Representing Haar wavelet transform, channel segmentation, and Convolutional operations, spatial attention mechanisms, channel attention mechanisms, and upsampling operations; , , , These represent the low-frequency approximate components, horizontal components, vertical components, and diagonal components obtained through Haar wavelet transform, respectively. , , These are the query, key, and value obtained after introducing a self-attention mechanism for the low-frequency approximate component, respectively. , , These are the query weighted results, key weighted results, and value sum results, respectively.

8. The feature extraction and dynamic fusion system for remote sensing small target detection according to claim 6, characterized in that: The local branch is a multi-scale depthwise separable convolutional parallel architecture, which includes three parallel branches. The three branches are depthwise separable convolutional branches with 3×3 convolutional kernels, 5×5 convolutional kernels, and 7×7 convolutional kernels, respectively. The mathematical expression for the local branch is: ； In the above formula, , These represent the input and output features of the local branch, respectively. , , These represent depthwise separable convolutions with 3×3, 5×5, and 7×7 kernels, respectively.

9. The feature extraction and dynamic fusion system for remote sensing small target detection according to claim 6, characterized in that: The dynamic feature fusion module includes point convolution, RFAConv convolution, spatial shift, and an adaptive weighting mechanism. The dynamic feature fusion module performs dynamic feature fusion according to the following steps: first, it expands the channel dimension of the input features by point convolution to obtain an expanded feature map; then, it segments the expanded feature map into features according to the channel dimension. , , , for features Feature extraction is performed using convolutional RFAConv, and the extracted features are then compared with the features. The first intermediate feature is obtained by weighted fusion using an adaptive weighting mechanism. , the first intermediate feature ,feature Then, feature extraction is performed using RFAConv convolution, and the extracted feature results are concatenated by channels to obtain the target's external contour features. ; target external contour features With features The second intermediate feature is obtained by weighted fusion using an adaptive weighting mechanism. ; for the second intermediate feature First, positional information is injected through a spatial displacement module, and then point convolution is used to process the second intermediate features. Perform fine-grained texture feature extraction; combine the extracted texture features with the relevant data. The internal texture features of the target are obtained by weighted fusion using an adaptive weighting mechanism. Finally, the target's external contour features are... With target internal texture features Channel splicing is performed to obtain the output features of the dynamic feature fusion module.

10. The feature extraction and dynamic fusion system for remote sensing small target detection according to claim 9, characterized in that: The mathematical expression for the dynamic feature fusion module is: ； ； ； ； ； ； In the above formula, , These represent the input and output features of the dynamic feature fusion module, respectively. , , , These represent channel segmentation, pointwise convolution, channel concatenation, and convolution RFAConv, respectively. For spatial displacement operations; , , , , , These are the dynamic weights of the corresponding features; , These are the first intermediate feature and the second intermediate feature, respectively. , These are the target's external contour features and internal texture features extracted by the dynamic feature fusion module, respectively.

Images