Deep Learning Innovations for Underwater Waste Detection: An In-Depth Analysis

Underwater waste detection is a critical challenge for preserving aquatic ecosystems, particularly due to inherent underwater distortions such as light refraction, occlusion, and scattering. In this study, we present a novel deep learning framework for real-time underwater waste detection by evaluating state-of-the-art object detection algorithms on a manually annotated custom dataset comprising 10,000 curated augmented images across various water bodies to simulate real-world turbidity, illumination, and occlusion, consisting of three classes: Underwater Trash, Rover, and Biological Life. Our approach incorporates robust feature extraction and specialized data augmentation techniques that effectively mitigate the adverse effects of underwater distortions. We investigate multiple architectures, including YOLOv8n, YOLOv7, YOLOv6s, YOLOv5s, Faster R-CNN, and Mask R-CNN, analyzing key parameters such as mean average precision (mAP), inference speed, and computational efficiency. Our results demonstrate that YOLO-based models achieve faster & accurate performance, with YOLOv8n and YOLOv7 reaching a detection accuracy of 96%, compared to 81% and 83% for Faster R-CNN and Mask R-CNN, respectively. Computationally, YOLOv8n achieves 76 ms per frame, while YOLOv5s runs at 14 ms per frame, confirming real-time viability for AUV deployment. The proposed approach offers significant advantages over existing methods by enabling rapid, accurate detection with low computational overhead, thereby paving the way for integration with autonomous underwater vehicles (AUVs) for environmental monitoring and waste management.