YOLOv13-LSA: An Enhanced Target Detection Algorithm for Water Surface and Underwater Garbage with Multi-Stage Transfer Learning Framework

Abstract

Water-surface and underwater garbage detection remains challenging because aquatic images often contain low contrast, light attenuation, reflection, wave interference, blurred boundaries, and large scale variation. These factors cause missed detections of small debris and inaccurate localization of irregular objects. Although YOLO-based methods have improved aquatic detection, many focus on a single direction, such as lightweight design, attention enhancement, or loss optimization, and still struggle with boundary ambiguity, multi-scale conflict, and cross-domain shift. To address these issues, this study develops YOLOv13-LSA by integrating existing SlideLoss, LSKA, and ASFF components into YOLOv13. SlideLoss is adopted to strengthen boundary regression for irregular and elongated targets, LSKA is embedded to enhance weak and low-contrast features, and ASFF is incorporated for adaptive multi-scale fusion. A source-to-target aquatic transfer learning framework is further designed by pretraining on underwater garbage data and fine-tuning on water-surface garbage data. Experiments show that YOLOv13-LSA improves detection accuracy, small-target recall, and robustness in complex aquatic scenes. Transfer learning further increases mAP@50 to 83.46% and reduces training time by 7.57%, supporting intelligent aquatic garbage monitoring.