A Reflection-Robust and Lightweight YOLOv8-Based Model for Water-Surface Plastic Bottle Detection

Abstract

With the increasing problem of water surface pollution, floating debris—especially plastic bottles—has become one of the main pollutants in inland rivers and lakes. Accurate and real-time detection of such floating objects is crucial for the autonomous operation of unmanned cleaning vessels (USVs). However, reflection, refraction, and wave interference on the water surface often cause severe false detections and missed detections. To address these challenges, this paper proposes an improved YOLOv8-based detection algorithm for plastic bottles on water surfaces. The model introduces a physics-prior feature extraction module (SIDSFront) to suppress specular highlights through reflection-invariant projection and dual-branch gate fusion, achieving illumination-insensitive feature representation. Furthermore, a shallow P2 detection layer is added to enhance small-object perception, and a wP2 + wHL weighted fusion strategy is designed to adaptively integrate multi-scale features while dynamically suppressing highlight regions. Experimental results on the FloW, IWHR_AI_Lable_Floater_ V1, and FloatingTrash datasets demonstrate that the proposed model improves mAP50 by 5.02% and mAP50–95 by up to 2.47% compared to the baseline YOLOv8, while maintaining over 80 FPS inference speed on RTX 5060 Ti. The method balances detection precision, robustness, and real-time performance, providing a reliable perception solution for intelligent unmanned cleaning vessels.