RG-YOLO: multi-scale feature learning for underwater target detection

Underwater target detection is pivotal in advancing marine development. However, the intricate underwater environments pose substantial challenges to this task, including dense target distributions, low contrast, and limited lighting conditions. To improve the performance of underwater target detection, this paper proposes an underwater target detection model based on YOLOv7-tiny framework, named RG-YOLO. Firstly, an efficient reparameterized multi-scale fusion (RMF) module is designed to boost the feature extraction capability of the model, capturing both local and global information and learning richer multi-scale feature representations. Secondly, this paper proposes a gather and distribute feature pyramid network (GDFPN) to significantly enhance the ability of multi-scale feature fusion. In addition, a dynamic head module is incorporated into the network. This module elevates detection performance for underwater small targets by unifying object detection heads with attentions. Finally, Shape-IoU is introduced to emphasize the inherent attributes of the bounding boxes, thereby obtaining higher accuracy in localization. Experimental results on the UTDAC2020, DUO, and RUOD datasets demonstrate that RG-YOLO achieves mean average precision (mAP) @50 of 85.1%, 86.1%, and 85.7%, respectively. These results represent improvements of 2.9%, 1.6%, and 2.3% over the baseline model, showing superior performance in underwater target detection compared to other state-of-the-art models.