An Improved VFNet Network for Ship Detection in SAR Imagery

Accurate detection of ships in high-resolution synthetic aperture radar (SAR) imagery plays an important role in effective maritime search and rescue, as well as maritime surveillance and monitoring. Recently, many deep learning-based detection networks have been proposed for SAR ship detection. However, there are still some detection challenges, particularly in inshore regions where the complex background interference from man-made structures frequently lead to a high false alarm rate. In addition, small ships with limited pixels are easily to be missed detection, especially in inshore regions. To solve the aforementioned problems, we develop an anchor-free SAR ship detection method based on improved VarifocalNet (VFNet). Specifically, a refined channel attention and fusion module (RCAFM) is designed to improve the detection performance. Within this module, the refined channel attention module is utilized to suppress the interference of irrelevant background information and enhance the detection network attention on the areas related to ship detection. Furthermore, the refined feature map is integrated with semantically richer feature map to improve the detection performance of ship in SAR imagery. The experiments conducted on the HRSID dataset demonstrate that the proposed method achieves an average precision of $92.3 \%$ at an IoU threshold of 0.5 and $80.0 \%$ at an IoU threshold of 0.75, which outperforms the baseline by $0.9 \%$ and $1.2 \%$, respectively.