Abstract
Accurate detection of Anthropogenic Floating Debris (AFD) is crucial for water pollution management. However, existing detection algorithms often lack environmental robustness in complex aquatic environments, and their high computational costs also impede deployment on lightweight platforms. To address these issues, this study proposes BiDB-YOLOv8, with an enhanced feature processing architecture. The model first employs a multi-branch convolutional block to improve the quality of extracted features, enhancing its ability to distinguish small targets from noisy backgrounds. It then uses an efficient bidirectional fusion network to ensure these high-quality features are effectively integrated across different scales. Additionally, this paper constructs and releases an original dataset, Turbid-floater, which includes diverse interference scenarios. Experimental results show that BiDB-YOLOv8 improves detection accuracy over the baseline YOLOv8n with a negligible increase in parameters. Compared to the larger YOLOv8s, it achieves nearly identical accuracy with only a fraction of the computational cost, highlighting a favorable trade-off between precision, robustness, and efficiency for real-world environmental monitoring.