Harnessing Deep Learning for Underwater plastic Trash Identification
- Version
- Download 47
- File Size 931.77 KB
- File Count 1
- Create Date 23 April 2024
- Last Updated 23 April 2024
Harnessing Deep Learning for Underwater plastic Trash Identification
Authors: Pradyumna K, Ravikumar A, Pancham H N, Palaniswamy I
Abstract—Machine and deep learning (DL) offer significant opportunities for exploring and monitoring oceans and for tackling important problems ranging from litter and oil spill detection to marine biodiversity estimation. Reasonably priced hardware platforms, in the form of autonomous (AUV) and remote operated (ROV) underwater vehicles, are also becoming available, fuelling the growth of data and offering new types of ap- plication areas. This article presents a research vision for DL in the oceans, collating applications and use cases, identifying opportunities, constraints, and open research challenges. We conduct experiments on underwater marine litter detection to demon- strate the benefits DL can bring to underwater envi- ronments. Our results show that integrating DL in underwater explorations can automate and scale-up monitoring, and highlight practical challenges in enabling underwater operations.
This project introduces a refined YOLOv8-based algorithm tailored for the en- hanced detection of small-scale underwater debris, to mitigate the prevalent challenges of high miss and false detection rates . The research presents the YOLOv8 algorithm, which optimizes the backbone, neck layer, and C2f module for underwater characteristics and incorporates an effective attention mech- anism. This algorithm improves the accuracy of underwater trash detection while simplifying the computational model. Empirical evidence underscores the superiority of this method over the other conventional network, manifesting in a significant uplift in detection performance. Notably, the proposed method realized a 63% mean average precision (mAP50), a 60% surge in recall (R). Transcending its foundational utility in marine conservation, this methodology harbors potential for subsequent integra- tion into remote sensing ventures. Such an adaptation could substantially enhance the precision of detection models, particularly in the realm of localized surveillance, thereby broad- ening the scope of its applicability and impact.