YOLO-MES: An Effective Lightweight Underwater Garbage Detection Scheme for Marine Ecosystems

Abstract

Marine pollution significantly impacts the sustainable development of marine ecosystems and the marine economy. Accurate detection of marine debris is essential for effective pollution control. However, existing high-precision detection algorithms are challenging to deploy on performance-constrained IoT underwater devices due to their large computational complexity and model size. To address this issue, this paper introduces YOLO-MES, an innovative underwater debris detection algorithm that integrates a lightweight design and a feature enhancement mechanism to enable efficient model deployment while preserving high accuracy. Unlike traditional YOLO models, YOLO-MES incorporates MobileNetV3 into underwater target detection, replacing the CSPDarknet backbone network, and optimizes the C3 and Conv layers through the bneck structure, significantly reducing computational demands and parameter scale at the architectural level. Additionally, YOLO-MES embeds the Efficient Channel Attention (ECA) module within the bneck structure to form the MECAneck module, which enhances adaptive feature extraction, significantly improving the network’s cross-channel feature capture and target recognition capabilities. This paper also proposes a streamlined Slim-neck design strategy, which effectively reduces the number of parameters in the neck network while maintaining multi-scale feature fusion accuracy. Experimental results indicate that YOLO-MES achieves 95.8% accuracy on the dataset, while reducing model size and computational complexity by 64% and 67%, respectively. Compared to existing mainstream detection algorithms, YOLO-MES offers significant advantages in lightweight design and computational efficiency, providing a practical and deployable solution for underwater target detection on mobile devices.

Keywords

• Deep learning
• underwater garbage detection
• lightweight
• MobileNetv3