Scene Characteristic Mining-Based Semisupervised Network for Ship Detection in SAR Images

Deep synthetic aperture radar (SAR) ship detection networks can achieve excellent performance by training with lots of target-level labels. However, obtaining target-level labels is extremely difficult in practice. In this paper, a scene characteristic mining-based semi-supervised network (SCMS-Net) is proposed for SAR ship detection to address this issue. This network can leverage scene-level labels of SAR images for enhanced performance when the target-level labels are limited. Our approach extends the fully supervised RefineDet network by incorporating a dedicated scene characteristic mining branch (SCMB). By effectively utilizing scene-level labels, SCMB can extract scene characteristics in SAR images that are beneficial for representing ships and clutter, thereby achieving higher performance. Additionally, a hierarchical test process is proposed in this paper. It can reduce false alarms in inland and inshore scenes, as well as missing ships in the offshore scene. Experiments based on an authoritative SAR ship detection dataset validate the effectiveness of our method.