Smart Waste Underwater Segmentation: Deep Learning Based Approach

Abstract

Waste sorting is a crucial process for efficient waste management, with automation being a significant factor for waste companies. Advancements in robotics, artificial intelligence, and autonomous driving have enabled underwater waste cleaning robots to accurately locate and recognize underwater waste. The image segmentation method, compared to deep-learning based target detection, provides a more refined and accurate approach to target recognition, improving environmental perception and efficiency in underwater waste sorting. To simplify the process, we proposed two trained models (YOLOv8and Mask RCNN). We used them to segment underwater waste and then facilitate the process of sorting the waste into different types such as mask, glass bottles, plastic bottles, electronics, metal, tire, plastic bags and waste. The proposed system is tested on Underwater_debris dataset. Our YOLOv8 model achieved 84% of average precision after 200 training epochs and an average precison that exceeds 83.3% for our Mask RCNN model after 100 training epochs. The underwater waste separation process is made and more intelligent by our proposed waste segmentation approach.