SAR small target detection method based on coordinate perception attention and spatial semantic context

Abstract

The application discloses a SAR small target detection network based on coordinate perception attention and spatial semantic context. Target detection is an important task of remote sensing image intelligent processing. Through a series of algorithms, the target on the image can be automatically positioned and recognized. Due to the imaging mechanism of SAR, there are many speckle noise points in the image, which makes it difficult to distinguish the target from the background, so compared with optical remote sensing images, the processing of SAR images is more difficult. The application can significantly improve the detection effect of small-scale targets of SAR remote sensing images, especially for the SAR remote sensing small target detection scene with complex background and serious noise interference, on the basis of effectively suppressing background noise, the detection rate and detection precision are significantly improved, the robustness and generalization ability of the model are strengthened, and the adaptability of the model to different scenes is improved.

Description

Technical Field

[0001] This invention relates to the field of small-scale target detection technology in SAR remote sensing image processing, specifically to a method for small target detection in SAR remote sensing images based on coordinate-aware attention and spatial semantic context. It can achieve good target detection results in SAR remote sensing scenes with complex backgrounds and severe noise interference; the feature representation of small targets is enriched, and the recognition accuracy is significantly improved. Background Technology

[0002] Synthetic Aperture Radar (SAR), due to its all-weather, all-day imaging capabilities, has been widely applied in military and civilian fields in recent years, with the rapid development of airborne and satellite-based systems. Applications include maritime traffic control, fisheries management, and maritime emergency rescue. Target detection is a crucial task in intelligent remote sensing image processing. Through a series of algorithms, targets in images are automatically located and identified. Traditional SAR image detection algorithms are mainly based on Constant False Alarm Rate (CFAR) detectors, which adaptively calculate the detection threshold by estimating the statistics of background clutter and maintaining a constant false alarm rate. However, these traditional methods are cumbersome to design manually, computationally complex, and have weak generalization ability, limiting their application in transfer learning. Furthermore, these traditional methods require a high level of expertise from researchers and are prone to overfitting.

[0003] In recent years, the rapid development of deep learning has significantly improved object detection performance. Convolutional neural network (CNN)-based methods can adaptively learn high-level semantic representations of images, completing object classification and localization through powerful training algorithms. Due to their excellent feature extraction and representation capabilities, CNNs have become the mainstream algorithm for object detection. Based on their model framework, mainstream object detection networks can be divided into two main categories: single-stage detection networks represented by the YOLO series and two-stage detection networks represented by Faster R-CNN. Two-stage networks prioritize detection accuracy, while single-stage networks prioritize detection efficiency.

[0004] Due to the unique imaging mechanism of SAR, images contain numerous speckle noise points, making it difficult to distinguish targets from the background. Therefore, SAR images are more challenging to process than optical remote sensing images. Furthermore, limited resolution results in small target scales in SAR images, lacking sufficient appearance features, making it difficult for detectors to extract discriminative information and hindering accurate target localization and identification. Inspired by human selective attention, some studies have used channel-space hybrid attention mechanisms to enhance the model's attention to foreground regions. However, these methods neglect spatial location information in channel attention, leading to spatial misalignment between channel-space attention features and spatial attention features, limiting the accuracy improvement of small target detection. Other works attempt to mine contextual relationships, utilizing environmental information to enrich the feature representation of small targets. However, due to the lack of effective fusion of local and global context, detectors struggle to fully utilize the essential relationship between targets and the environment. Summary of the Invention

[0005] In view of this, the present invention provides a small target detection method for SAR remote sensing images based on coordinate-aware attention and spatial semantic context. This method first uses YOLOX as the basic framework for target detection, introducing a coordinate-aware attention mechanism into the network feature extraction module. By extracting and encoding features in the horizontal and vertical directions respectively, spatial coordinate information is embedded into channel attention, allowing the model to maintain the spatial relationship of features while acquiring channel attention. Coordinate-aware channel attention is used to enhance the input features, and spatial attention is further mined on the enhanced feature map, ensuring that the spatial relationship between spatial attention features and channel attention features remains aligned, resulting in a target feature map enhanced by coordinate-aware hybrid attention. After passing through multiple stacked coordinate-aware attention-enhanced feature extraction modules, a series of multi-scale, multi-level coordinate-aware hybrid attention-enhanced feature maps are obtained. Next, some of the extracted feature maps are input into the spatial semantic context module. On the one hand, multi-scale spatial context features are obtained by extracting local spatial environment features within different receptive fields of the target; on the other hand, the global semantic context association between the target and the entire scene is obtained through the summarization and interaction of global features. The multi-scale spatial context and the global semantic context are fused to obtain a spatial semantic context feature map. Then, the multi-scale coordinate-aware attention-enhanced feature map and the spatial semantic context feature map are fed into the YOLOX feature fusion network for feature fusion. Through a top-down and bottom-up fusion path, a multi-scale target fusion feature map is obtained. Finally, the multi-scale target fusion feature map is input into the target bounding box detection head to obtain the target detection result.

[0006] Beneficial effects:

[0007] (1) The present invention designs a coordinate-aware attention mechanism that encodes spatial coordinate information into channel attention and further mines spatial attention features on the features enhanced by coordinate-aware channel attention, so that the spatial position distribution of channel attention features and spatial attention features remains aligned, which alleviates the spatial misalignment problem of the previous channel-space hybrid attention mechanism, refines the granularity of the attention mechanism, strengthens the ability to focus on small targets during feature extraction, improves the positioning accuracy of small targets, and suppresses the interference of background noise.

[0008] (2) The present invention designs a spatial semantic context method, which extracts local environmental features in different spatial ranges around the target and constructs global semantic association between the target and the whole scene, so that the target can show richer spatial and semantic features, thus making it more recognizable and improving the detection performance of the model for small targets.

[0009] (3) The present invention can significantly improve the detection effect of small-scale targets in SAR remote sensing images. In particular, for SAR remote sensing small target detection scenarios with complex backgrounds and severe noise interference, it can significantly improve the detection rate and detection accuracy on the basis of effectively suppressing background noise, and enhance the robustness and generalization ability of the model, thereby improving the model's adaptability to different scenarios. Attached Figure Description

[0010] Figure 1 For small target detection network structure based on coordinate-aware attention and spatial semantic context Detailed Implementation

[0011] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0012] This invention extracts spatially aligned coordinate-aware hybrid attention-enhanced feature maps through a coordinate-aware attention mechanism, enhancing the model's ability to focus on small targets, improving the localization accuracy of small targets, and suppressing background noise interference. Through a spatial semantic context method, it fully mines and combines local spatial information and global semantic relationships of the environment, enriching the spatial semantic feature expression of the target and improving target recognition. The model structure and details of each method module are as follows... Figure 1 As shown.

[0013] Step 1: Multi-scale coordinate perception and hybrid attention-enhanced feature extraction

[0014] This invention proposes a Coordinate-Aware Attention CrossStage Partial (CAA-CSP) feature extraction module. The input image is first scaled by the Focus module of the YOLOX network, and then subjected to several consecutive Coordinate-Aware Attention Feature Extraction modules for feature extraction from shallow to deep, resulting in a multi-scale, multi-level spatially aligned coordinate-aware hybrid attention-enhanced feature map.

[0015] Specifically, it includes:

[0016] The input features are first downsampled by a 3×3 convolution with a stride of 2. Then, they are processed by two separate 1×1 convolutions, halving the feature dimensions and feeding them into two different processing paths. Feature map A from one path is first reduced in channel count by a 1×1 convolution, then processed by a 3×3 depthwise convolution before entering the Coordinate-Aware Attention (CAA) module to obtain a coordinate-aware hybrid attention-enhanced feature map. Next, a 1×1 convolution restores the channel count, and finally, the resulting feature map is element-wise added to feature map A. The resulting feature map is then stacked channel-wise with feature map B from the other path, and finally, a 1×1 convolution refines the merged feature map.

[0017] In the coordinate-aware attention module, coordinate-aware channel attention is first calculated. Specifically, the input feature map is first subjected to one-dimensional average pooling on both the X and Y axes, resulting in two two-dimensional feature maps representing coordinate-aware inductive feature maps in the horizontal and vertical directions. These two feature maps are then fed into two independent sets of convolution operations, each consisting of two 1×1 convolution operations and a sigmoid activation function. The two resulting feature maps are then multiplied to obtain the coordinate-aware channel attention map, which is then multiplied with the input feature map to obtain the coordinate-aware channel attention-enhanced feature map. Next, spatial attention is further mined based on the coordinate-aware channel attention-enhanced feature map. Specifically, the coordinate-aware channel attention-enhanced feature map is first subjected to max pooling and average pooling operations in the channel direction, resulting in two spatial inductive feature maps. These two feature maps are then stacked along the channel direction, and the resulting feature map is sequentially fed into two 3×3 convolutions and a sigmoid activation function to obtain the coordinate-aware spatial attention map. Finally, the coordinate-aware spatial attention map is multiplied with the coordinate-aware channel attention-enhanced feature map to obtain the coordinate-aware hybrid attention-enhanced feature map.

[0018] Step 2: Spatial Semantic Context Extraction

[0019] This invention proposes a Spatial-Semantic Context (SSC) module. Partial multi-scale, multi-level coordinate-aware hybrid attention-enhanced feature maps are fed into the SSC module, simultaneously capturing local spatial context information of the target and its environment, as well as global semantic context associations, to obtain a spatial semantic context feature map.

[0020] Specifically, it includes:

[0021] The input feature map is first fed into three parallel deep-dilated convolutional operations. Based on different kernel sizes (3×3, 3×3, 5×5) and dilation rates (1, 3, 3), three sets of local spatial feature maps with different receptive fields are obtained. These three sets of feature maps are then interleaved according to their corresponding channels, and a 1×1 group convolution is used to fuse every three consecutive local spatial feature maps into a single feature map. Next, a sigmoid activation function is applied to obtain a multi-scale spatial context feature map. Simultaneously, the input feature map undergoes a global pooling operation, two 1×1 convolutions, and a sigmoid activation function, and the result is multiplied by itself to obtain a global semantic context feature map. Finally, the multi-scale spatial context feature map and the global semantic context feature map are added together to obtain a spatial semantic enhancement feature map.

[0022] Step 3: Multi-level feature fusion

[0023] This invention utilizes the YOLOX feature fusion network to further interact and refine location and semantic information by combining coordinate-aware hybrid attention-enhanced feature maps and spatial semantic context feature maps through bottom-up and top-down cross-level feature fusion paths, resulting in multi-scale target fusion feature maps.

[0024] Step 4: Output of Target Detection Results

[0025] This invention utilizes the YOLOX target bounding box detection head to organize and summarize the multi-scale target fusion feature map, and finally outputs the target classification feature map, the target bounding box position regression feature map, and the target bounding box confidence regression map. After a series of post-processing steps, the target's position and category inference results are obtained.

Claims

1. A SAR small target detection method based on coordinate-aware attention and spatial semantic context, comprising the following steps: Step one, multi-scale coordinate-aware hybrid attention enhanced feature extraction: The SAR image is first input into the Focus module of the YOLOX network for image scaling; then the scaled features are input into several consecutive coordinate-aware attention feature extraction modules for feature extraction from shallow to deep, obtaining multi-scale multi-level spatially aligned coordinate-aware hybrid attention enhanced feature maps; Specifically, the input features are first subjected to a 3x3 convolution operation with a stride of 2 for two times of downsampling, and then passed through two convolution kernels with a size of 1x1 for convolution calculation, reducing the feature dimension by half while being sent to two different processing paths respectively; the feature map A on one of the paths is first reduced in channel number by a 1x1 convolution, then processed by a 3x3 deep convolution, and then input into the coordinate-aware attention module to obtain the coordinate-aware hybrid attention enhanced feature map; then the channel number is restored by a 1x1 convolution, and finally the obtained feature map is added element by element with feature map A; the obtained feature map is stacked with feature map B on the channel, and finally a 1x1 convolution is performed to refine the merged feature map; In the coordinate-aware attention module, first, coordinate-aware channel attention calculation is performed, specifically, the input feature map is first subjected to one-dimensional average pooling on the X and Y axes to obtain two two-dimensional feature maps, which represent the coordinate-aware inductive feature maps in the horizontal and vertical directions; the two feature maps are sent to two independent convolution operations respectively, each operation includes two 1x1 convolution operations and a Sigmoid activation function; then the obtained two feature maps are multiplied to obtain the coordinate-aware channel attention map, which is then multiplied with the input feature to obtain the coordinate-aware channel attention enhanced feature map; then the spatial attention is further mined based on the coordinate-aware channel attention enhanced feature map, specifically, the coordinate-aware channel attention enhanced feature map is first subjected to maximum pooling and average pooling operations in the channel direction to obtain two spatial inductive feature maps; the two feature maps are stacked on the channel to obtain a feature map which is then sent to two 3x3 convolutions and a Sigmoid activation function in turn to obtain the coordinate-aware spatial attention map; finally, the coordinate-aware spatial attention map is multiplied with the coordinate-aware channel attention enhanced feature map to obtain the coordinate-aware hybrid attention enhanced feature map; Step two, spatial semantic context extraction: The multi-scale multi-level spatially aligned coordinate-aware hybrid attention enhanced feature map obtained in step one is input into the spatial semantic context module to simultaneously capture the local spatial context information and global semantic context association of the target and the environment, obtaining a spatial semantic enhanced feature map; Specifically, the input feature map is first input into three parallel deep dilated convolution operation processes, based on different convolution kernel sizes (3x3, 3x3, 5x5) and dilation rates (1, 3, 3), three groups of local spatial feature maps with different size receptive fields are obtained; the three groups of feature maps obtained are interleaved in turn according to the corresponding channels, and then each continuous three local spatial feature maps are fused into one feature map through a 1x1 grouped convolution; then a Sigmoid activation function is used to obtain a multi-scale spatial context feature map; at the same time, the input feature map is subjected to a global pooling operation, two 1x1 convolutions and a Sigmoid activation function, and the result is multiplied by itself to obtain a global semantic context feature map; finally, the obtained multi-scale spatial context feature map and the global semantic context feature map are added to obtain a spatial semantic enhancement feature map; Step three, multi-level feature fusion: Using the feature fusion network of YOLOX, the multi-scale multi-level spatial alignment coordinate perception hybrid attention enhancement feature map obtained in step one and the spatial semantic enhancement feature map obtained in step two are fused through a bottom-up and top-down cross-level feature fusion path to realize the interaction and refinement of position information and semantic information, and a multi-scale target fusion feature map is obtained. Step four, target detection result output: The multi-scale target fusion feature map obtained in step three is input into the target box detection head to output the detection result of the SAR small target.

2. The SAR small target detection method based on coordinate-aware attention and spatial semantic context according to claim 1, wherein, In step four, the target box detection head of YOLOX is used to arrange and summarize the multi-scale target fusion feature map, and finally output the target classification feature map, target box position regression feature map and target box confidence regression map, and then a series of post-processing is performed to obtain the position and category reasoning result of the target.

Images