Advancing Underwater Trash Detection: Harnessing Mask R-CNN, YOLOv8, EfficientDet-D0 and YOLACT

Every year, millions of tons of trash and other pollutants are introduced into the ocean. Marine animals and coral reefs face significant threats to their lives due to human negligence in properly disposing of trash. Millions of marine life perish every single year due to the contamination of water bodies. Before embarking on this research, several challenges were faced, including the scarcity of high-quality underwater trash datasets, the complexity of underwater environments, and the limitations of existing detection models in terms of accuracy and speed. Additionally, selecting suitable deep learning models presented significant hurdles due to the varying availability of pretrained models, differing capabilities in handling underwater imagery, and the trade-offs between detection accuracy and computational efficiency. Amid these increasing environmental concerns, this research explores novel deep learning models for detecting underwater trash namely Mask R-CNN, YOLOv8, EfficientDet-DO, and YOLACT, aiming to contribute to cleaner water bodies. These models were trained on the benchmark TrashCan 1.0 dataset. The pretrained weights of the COCO 2017 dataset were used for transfer learning. The study found that YOLOv8 has outperformed all the remaining models in terms of mAP, Recall, and F1 Score. YOLOv8 achieved the highest mAP (0.714), outperforming Mask R-CNN (0.627), EfficientDet-DO (0.560), and YOLACT (0.542). Notably, YOLOv8 trained 3 to 10 times faster than other models. This accelerated training time positions YOLOv8 as the preferred choice for real-world deployments of Autonomous Underwater Vehicles(AUVs) where rapid adaptation to new types of trash is essential. On the other hand, Mask R-CNN performs well at detecting small trash in complex underwater scenes. This detailed analysis can be used to improve the training data for YOLOv8, potentially boosting its accuracy in detecting smaller objects and complex under-water environments while maintaining its real-time processing advantage. This comprehensive analysis paves way for further advancements in trash detection algorithms, fostering a more effective mitigation strategy against ocean pollution.