Marine Debris detection and classification using Deep Learning Techniques

Abstract

Introduction: Ocean health faces an increasing danger from marine debris because such pollution endangers marine ecosystems and threatens marine life populations. Worldwide waste production continues its upward trend while many materials remain in the environment throughout multiple years thus surpassing current mitigation solutions. Multiple challenges exist when trying to assess debris from different regions because assessment methods lack standardization across timeframes. Deep learning has proven itself as an advanced automated system for detecting and categorizing marine debris in current conditions. This research evaluates the performance of modern object detection systems Faster R-CNN, SSD, and YOLOv5. We introduce a novel confidence adjustment technique tailored for SSD, designed to improve detection reliability under challenging underwater conditions.

Objectives:

(1) Evaluate the performance of three deep learning-based models (YOLOv5, Faster R-CNN, and SSD) in localizing and classifying marine debris from images and videos,

(2) Propose a confidence adjustment method for the SSD model in order to increase the detection rate of marine debris in underwater environments.

(3) Compare the performance of the detection models on the JAMSTEC dataset in terms of localization accuracy, precision, recall, and mean average precision (mAP).

Methods: The study conducts a comparative analysis between YOLOv5 , SSD with adjusted confidence and Faster R-CNN for object detection on marine debris images.

Results:YOLOv5 performed the best with the highest score of 91% on mAP@0.5, and it also had the best performance on small and partially hidden objects. Faster R-CNN was able to detect marine debris, and it was able to detect marine debris in poor water quality, therefore the score was 87% on mAP@0.5. SSD performed better after the confidence score fix, thus the score was 84% on mAP@0.5. In particular, it was best at identifying small and unusual trash such as small plastics, because this indicates that the new confidence correction is very useful for SSD. It makes the SSD much better at detecting trash in the ocean immediately, so it is a significant improvement for ocean trash detection.

Conclusions: Each hindrance within the research process gave vital lessons which collectively produced the better structured methodology that emerged in this work. This investigation focused on evaluating efficient deep learning techniques for detecting important marine debris objects. An evaluation of Faster R-CNN, SSD, and YOLOv5 detection models helped identify the fundamental characteristics as well as the operational boundaries of each system. YOLOv5 proved to be the most accurate model but adding novel confidence adjustment techniques to SSD produced better performance results demonstrating that careful improvements work well for lightweight systems.

Further evaluations using multiple datasets such as DeepTrash and TrashCAN are planned to validate model generalizability. Additionally, incorporating cross-validation and domain adaptation techniques may strengthen model robustness in new environments. Finally, we aim to integrate our confidence-adjusted SSD model into low-power embedded systems for field testing.