AAYOLO: Robust Detection of Small and Occluded Floating Objects in Water Bodies Under Varying Illumination Using the FloW Dataset

Abstract

Floating waste detection is a critical task for environmental monitoring and preservation, particularly in water bodies polluted by plastic waste and other debris. Traditional detection models like YOLOv5s struggle with detecting small, occluded, and poorly illuminated floating objects due to suboptimal feature extraction and limited datasets. This study proposes an enhanced model, AAYOLO (Atrous-Attention YOLO), which integrates atrous convolutions and enhanced channel attention (ECA) mechanisms to improve feature extraction across various scales. Atrous convolutions are applied at strategic levels of the network (P1/2, P3/8, and P5/32) to expand the receptive field, while ECA effectively adjusts feature weights to enhance object detection under varying lighting conditions. The proposed model is evaluated on the FloW-Img dataset, a benchmark specifically designed for floating waste detection, consisting of high-resolution images of plastic bottles found in aquatic environments. Experimental results demonstrate that AAYOLO outperforms the baseline YOLOv5s model, achieving a higher mean Average Precision (mAP@0.5) of 0.892 compared to 0.872, along with improvements in precision (0.878 vs. 0.858) and recall (0.836 vs. 0.820). This approach presents a promising solution for real-time, accurate detection of floating debris, contributing to improved waterway monitoring and environmental protection.