A Sidelobe-Aware Semi-Deformable Convolutional Ship Detection Network for Synthetic Aperture Radar Imagery

This paper focus on SAR ship detection, which has been widely used in tasks such as marine traffic, fisheries management, battlefield posture assessment and military target reconnaissance. One popular solution is to utilise deep learning algorithms in conjunction with SAR image object detection. However, due to the unique imaging characteristics of SAR, existing solutions usually do not distinguish well between ships and interference targets in complex backgrounds. In this paper, we address this problem by proposing a sidelobe-aware semi-deformable convolution that takes full advantage of the combination of both standard and deformable convolution to learn high-quality ship features without significantly increasing the computational complexity. Specifically, it makes the feature extraction location more closely fit to the ship shape, strengthens the extraction capability of the region of the ship target itself and the sidelobe features, while reducing the extraction of background information. The channel attention mechanism is then proposed to enhance the ship local detail information extraction. Experiments on two widely used datasets show that the proposed method outperforms the state-of-the-art methods, which is effective and efficient to improve SAR ship detection.