An underwater garbage detection algorithm based on REG-YOLO

Abstract

Underwater garbage detection plays a crucial role in maintaining the balance of aquatic ecosystems. To address the inefficiency and high cost of traditional manual debris retrieval, as well as the challenge of degraded detection accuracy in complex underwater environments, an innovative underwater waste detection algorithm based on the REG-YOLO network is proposed. First, the backbone is structurally reconfigured by introducing the RFAConv module. Equipped with multi-scale spatial awareness, this module adaptively adjusts the receptive field. It strengthens multi-scale feature extraction, effectively improving target detection accuracy. Second, an EMAttention mechanism is embedded at the end of the backbone to enhance the model's focus on critical features and its ability to model global context. This suppresses interference from complex underwater backgrounds, yielding more stable and accurate detection results. Finally, the GSConv module is integrated into the neck network to efficiently fuse and compress hierarchical features, which further boosts both recognition accuracy and inference efficiency. Experiments on a public marine debris dataset demonstrate that the proposed REG-YOLO achieves a precision of 81.3%, a recall of 75.0%, and mAP@50 of 81.0%, while using only 2.21M parameters and 5.7 GFLOPs. Compared to the baseline YOLO26n model, the proposed approach improves Precision by 2.5%, Recall by 8.2%, and mAP@50 by 2.3%, while simultaneously reducing the parameter count and computational load by 12.12% and 1.72%, respectively. Demonstrating significant advantages in both Precision and lightweight design, this algorithm provides a highly efficient and practical solution for underwater garbage detection, offering substantial value for protecting marine environments.