RED-YOLO: A lightweight synergistic fusion framework with channel-efficient neck and multi-dimensional attention for underwater small object detection

Abstract

Accurate object recognition in the marine environment is critical for protecting biodiversity and ensuring the sustainable exploitation of ocean resources. While state-of-the-art techniques based on deep learning have demonstrated impressive accuracy, their high computational requirements prevent real-time deployment, particularly on resource-constrained platforms such as AUVs and ROVs. To overcome this limitation, we present RED-YOLO, a lightweight yet accurate model tailored for underwater vision. Firstly, we introduced an Efficient Multi-scale Attention (EMA) module into the backbone to strengthen multi-resolution feature extraction. In addition, the neck structure incorporates a Lightweight-RepGFPN design to simplify feature fusion while reducing redundant parameters. Finally, we replaced DCNV2 with DCNV4, which employs a scale, spatial, and task-aware attention mechanism to facilitate target localization. Our model outperforms other models with just 2.8 M parameters and 9.7 GFLOPs, achieving 87.3 % mAP on our comprehensive test dataset. We believe that RED-YOLO provides a promising solution for underwater object detection by striking a balance between computational efficiency and detection accuracy.