Enhancing YOLO performance with attention module for plastic and non-plastic waste detection on water surfaces

Abstract

The rapid accumulation of plastic waste in aquatic environments poses serious threats to ecosystems, water management systems, and human health. This growing concern creates an urgent need for efficient and accurate detection methods. To address this challenge, this work proposes an approach to enhance YOLO performance by integrating attention modules for plastic and non-plastic waste detection on water surfaces. A comprehensive evaluation is conducted on the Plastic on Water dataset, considering detection accuracy, computational complexity, and inference speed. The results identify YOLO11n as the most effective baseline, achieving a mean Average Precision (mAP) of 96.3% with 2,590,230 parameters, 6.4 GFLOPs, and an inference speed of 18.58 FPS. To further improve performance, several attention modules are integrated into the YOLO11n architecture. Among them, the Convolutional Block Attention Module (CBAM) yields the best performance, achieving an mAP of 96.7% with 2,598,520 parameters and 6.5 GFLOPs, while maintaining real-time performance at 18.26 FPS. The results demonstrate improved detection capability, particularly for small and less prominent objects, with negligible additional computational cost. These findings highlight the effectiveness of attention mechanisms, especially CBAM, in enhancing lightweight object detection models for real-time aquatic waste monitoring.