Enhanced Vehicle Detection in SAR Images via Global and Dual Attention Mechanisms

Target detection in Synthetic Aperture Radar (SAR) images is widely applied in various fields, such as geographic information systems, environmental monitoring, disaster warning, urban planning, and military surveillance. Despite the advances in mainstream detection methods, challenges persist in handling complex backgrounds, small targets, and targets with arbitrary orientations. To address these issues, we propose the Global and Dual Attention Network (GDAN), which builds upon the YOLOv5 framework. GDAN incorporates the Spatial Pyramid Pooling Feature Concatenation and Spatial Pyramid Convolution (SPPFCSPC) module to process input images of various sizes and extract multi-level features. The Global Attention Mechanism (GAM) is designed to handle feature interactions in three dimensions simultaneously, enabling adaptive feature aggregation and improving the model's robustness. Furthermore, the Dual Attention (DA) module enables the model to receive multiple feature information from different spatial locations and adaptively acquire global image features. Experimental results demonstrate that GDAN achieves state-of-the-art performance with a mean Average Precision (mAP) of 82.63% on the HighSAR dataset and 91.5% on the VOC dataset, significantly outperforming existing methods. The links to the code https://github.com/ynlsj/Global-and-Dual-Attention-Mechanisms/, dataset(https://github.com/whu-csl/SAR_vehicle_detection_dataset and http://host.robots.ox.ac.uk/pascal/VOC/voc2012/ )