Small Target Ship Detection in SAR Images Using an Enhanced Lightweight YOLOv5 Model

Synthetic aperture radar (SAR) can generate high-resolution images in all weather, making it widely used for maritime ship detection, which plays a crucial role in maritime traffic management and safety. However, the abundance of noise and the small targets in SAR images pose challenges for ship detection. To detect ships more efficiently and accurately, this paper presents targeted improvements to you only look once-v5 (YOLOv5). First, we propose a depthwise separable C3 module (DSC3), which replaces traditional convolutions with depthwise separable convolutions, reducing the model's parameters and making the model more lightweight. Second, we propose an improved path aggregation network (IMPANet), which utilizes a shallow high-resolution feature as an additional detection head, effectively enhancing the model's ability to detect small targets. Additionally, it uses skip connections to maintain the global structural information of small ships. Finally, we propose a small ship feature extraction module (SSFE), which utilizes multi-scale convolutions and the proposed three-dimensional hybrid attention mechanism (THA) to fully extract features of small target ships. Experiments on a high-resolution SAR images dataset (HRSID) confirm the effectiveness and superiority of our model. Compared to YOLOv5, our model achieves a 0.022 increase in AP while reducing parameters by 0.92M.