Effective Small Ship Detection with Enhanced-YOLOv7

Small ship detection is widely used in marine environment monitoring, military applications and so on, and it has gained increasing attentions both in industry and academia. In this paper, we propose an effective small ship detection algorithm with enhanced-YOLOv7. Specifically, to reduce the feature loss of small ships and the impact of marine environment, we firstly design a small object-aware feature extraction module by considering both small-scale receptive fields and multi-branch residual structures. In addition, we propose a small object-friendly scale-insensitive regression scheme, to strengthen the contributions of both bounding box distance and difficult samples on regression loss as well as further increase learning efficiency of small ship detection. Moreover, based on the formulated penalty model, we design a geometric constraint-based Non-Maximum Suppression (NMS) method, to effectively decrease small ship detection omission rate. Finally, extensive experiments are implemented, and corresponding results confirm the effectiveness of the proposed algorithm.