关键词: 水下目标检测, 特征融合, 注意力机制, YOLOv7, 计算机视觉

Abstract: Blurred underwater images and complex underwater environments lead to issues of poor performance and low detection accuracy of existing underwater target detection models. To address these issues, an improved YOLOv7 target detection model, YOLOv7-RMC, is proposed. Firstly, the residual fusion global attention mechanism (RGAM) is designed to extract more critical information from blurred images. It can address the loss of contextual feature information as the network layers deepen, thereby strengthening the model’s feature extraction capabilities. Secondly, to address the loss of information in both far and near targets, as well as the loss of defect information in small targets, the multi-scale shallow feature fusion network (MSFN) is introduced to optimize and reconstruct the original neck network structure. This network improves target detection accuracy through multiple interactions between deep and shallow features. Finally, a lightweight up-sampling operator module, the content aware reassembly of feature (CARAFE), is introduced into the optimized neck network MSFN. It enhances the network’s feature fusion capabilities and retains more target information through feature reorganization and expansion. Experimental results show that the proposed algorithm achieves mAP@0.5 of 87.4% and 97.9% on the URPC and Brackish datasets respectively. Compared to the original YOLOv7 network model, the mAP@0.5 is improved by 2.3 and 0.9 percentage points, demonstrating the improved detection capabilities and strong practicality of the proposed model.

Key words: underwater target detection, feature fusion, attention mechanism, YOLOv7, computer vision