Underwater Waste Detection Using Deep Learning A Performance Comparison of YOLOv7 to 10 and Faster RCNN

Technical Summary: Underwater Waste Detection Using Deep Learning

Problem Statement
Underwater pollution, particularly from plastics and chemical contaminants, poses severe threats to marine biodiversity, ecosystems, and human health. It disrupts food chains, damages coral reefs, and incurs significant economic losses in industries such as tourism and fishing. While traditional management strategies are essential, the detection and removal of underwater waste require advanced technological solutions. Current challenges include low visibility, light scattering, and particulate matter in underwater environments, which degrade image quality and hinder accurate waste identification. Although various machine learning approaches have been proposed, there is a lack of comprehensive comparative analysis among state-of-the-art object detection models specifically applied to multi-class underwater garbage detection.

Methodology
This study conducted a comparative evaluation of five cutting-edge object detection algorithms: YOLOv7, YOLOv8, YOLOv9, YOLOv10, and Faster R-CNN. The research utilized a dataset sourced from an online repository containing 5,130 underwater images categorized into 15 distinct classes of waste: masks, cans, cell phones, electronics, glass bottles, gloves, metal, net, polythene bags, plastic bottles, rods, sunglasses, tires, and miscellaneous items.

The dataset was split into training (3,628 images), validation (1,001 images), and testing (501 images) sets. All models were trained for 50 epochs with a batch size of 64 and an input image size of 128 pixels. The study employed standard evaluation metrics defined in the YOLO documentation:

• Precision: The proportion of genuine positives among all positive predictions.
• Recall: The proportion of true positives identified out of all actual positives.
• Mean Average Precision (mAP): The average precision value across all 15 object classes.

Key Contributions
The primary contribution of this work is a systematic performance comparison of the latest YOLO variants (v7 through v10) against the established Faster R-CNN architecture within the specific context of underwater waste detection. While previous studies often focused on optimizing a single model architecture (e.g., improving YOLOv5 or YOLOv8 for specific conditions), this study evaluates the relative efficacy of multiple modern architectures on a diverse, multi-class underwater dataset. The research identifies the most effective model for recognizing waste materials under challenging marine conditions without proposing new architectural modifications, relying instead on the inherent capabilities of the selected pre-trained models.

Results
The experimental results demonstrated that YOLOv8 outperformed all other models. The specific performance metrics were as follows:

• YOLOv8: Precision of 0.832, Recall of 0.727, and an mAP of 80.9%.
• YOLOv9: Precision of 0.789, Recall of 0.729, and an mAP of 79.0%.
• YOLOv10: Precision of 0.688, Recall of 0.686, and an mAP of 71.6%.
• YOLOv7: Precision of 0.663, Recall of 0.656, and an mAP of 67.9%.
• Faster R-CNN: Precision of 0.484, Recall of 0.629, and an mAP of 61.5%.

The study attributes YOLOv8's superior performance to its architecture, which incorporates advanced features such as improved anchor-free mechanisms and self-supervised learning capabilities, allowing for more precise recognition in varied settings. Confusion matrix analysis and metric curves for YOLOv8 confirmed its stability, with decreasing loss values and improving mAP and recall over training epochs. Visual predictions further validated the model's ability to detect various waste items accurately.

Significance and Limitations
The paper concludes that YOLOv8 is a reliable and effective tool for detecting a variety of underwater waste categories, offering improved detection capabilities and scalability for cleanup operations. The findings suggest that YOLOv8 is a strong candidate for integration into global efforts to mitigate underwater pollution.

However, the authors maintain a modest stance regarding the study's scope. They acknowledge limitations, noting that the dataset does not fully capture the complexity of real-world underwater conditions, such as extreme low visibility, varying illumination, and highly cluttered backgrounds. Additionally, the evaluation was limited to 15 specific classes, which may restrict the generalizability of the results to other waste types. The study does not claim to have solved all underwater detection challenges but rather identifies YOLOv8 as the most effective among the tested models for the specific dataset and conditions evaluated. Future work is suggested to involve larger, multi-location datasets and the integration of domain adaptation techniques to further enhance model generalization.