SFANet: Efficient Detection of Vehicle Targets in SAR Images Based on SAR-Specialized Feature Aggregation

With the swift progress of deep learning technologies, the detection of ground vehicles in synthetic aperture radar (SAR) images has emerged as a prominent research focus. In this article, research is conducted on the detection of ground vehicles using MiniSAR mounted on unmanned aerial vehicles. However, the substantial background interference and high similarity between ground vehicles make it challenging for object detection. To address these issues, an efficient object detection network based on SAR-specialized feature attention (SFA) and lightweight global detector (LGD) is proposed in this article. First, the SFA is proposed, which introduces proportional attention to different detection layers. By analyzing and focusing on spatial and channel information, it alleviates the classification challenges when dealing with highly similar ground vehicles and limited information. Second, a novel detector LGD is proposed, which combines Swin Transformer and lightweight convolutions. LGD captures deeper semantic information through global attention, enabling the network to detect vehicles in complex scenes. Finally, transfer learning is utilized to load pre-classification weights of the backbone. To verify the effectiveness of the proposed algorithm, comprehensive experiments were performed across multiple datasets. The results indicate that the proposed method offers substantial performance enhancements.