Improved YOLOv5s algorithm for maritime small target detection

Aiming at the problem such as missed detections and false detections in maritime small-target detection caused by the loss of small target features, an improved YOLOv5s-based algorithm named LSTD-YOLOv5s (leap small target detection- YOLOv5s) is proposed. In the multi-scale feature fusion framework, additional upsampling and downsampling operations are introduced to integrate shallower feature maps, along with an extra prediction layer to enhance semantic and positional information extraction. A Leap structure is added to preserve richer detail information. An efficient multi-scale attention mechanism is introduced to focus the model on key features. The SIoU loss is adopted to more accurately capture the positional relationships of targets of varying sizes within the image. Experimental results on the SeaDronesSee v2 dataset demonstrate that compared to YOLOv5s, the improved model achieves significant performance gains of 2.9% in recall, 2.0% in mAP50, and 1.5% in F1-score.