Detection of Floating Objects on Water Using Improved YOLOv5

The detection of floating objects on water surfaces holds paramount significance in the field of inland waterway environmental monitoring and management. However, the precise identification of small targets within images of floating objects poses a considerable challenge due to their diminutive size and limited textural features. In response, we propose an enhanced model for waterborne floating object detection based on the YOLOv5s architecture. This model incorporates the Squeeze-andExcitation (SE) attention mechanism to capture global information related to small targets and introduces the Contextual Transformer Module 3D (CT3) to improve the fusion of multiscale features. Evaluation on the publicly available Flow dataset demonstrates a noteworthy $2.7 \%$ enhancement in mean average precision at 0.5 intersection over union (mAP@0.5), indicating substantial progress in the accurate detection of small targets.