An Improved YOLOv8-Based Shallow Sea Creatures Object Detection Method

Abstract

With the development and utilization of marine resources, object detection in shallow sea environments becomes crucial. In real underwater environments, targets are often affected by motion blur or appear clustered, increasing detection difficulty. To address this problem, we propose an improved YOLOv8-based shallow sea creatures object detection method. We integrate receptive-field coordinate attention (RFCA) into the cross-stage partial bottleneck with the two convolutions (C2f) module of YOLOv8, creating the RFCA-enhanced C2f (C2f_RFCA). This enhancement improves feature extraction and fusion by leveraging multiscale receptive fields and refined feature fusion strategies, enabling better detection of blurred and occluded objects. The C2f_RFCA module captures both local and global features, enhancing detection accuracy in complex underwater scenarios. We additionally devised an improved dynamic head by substituting the deformable ConvNets version two (DCNv2) with DCNv3, forming dynamic head with DCNv3. This upgrade increases the flexibility of feature mapping and improves accuracy in detecting densely clustered objects by allowing adaptive receptive fields and enhancing boundary delineation. To evaluate the algorithm performance, we trained it on real-world underwater object detection data sets and conducted generalization experiments on detecting underwater objects, the underwater robot professional competition 2020 and underwater target detection and classification 2020 data sets. Experimental results show that, compared with YOLOv8n, our method increases mAP@0.5 by 1.9%, 1.7%, 4.3%, and 3.3%, and mAP@0.5:0.95 by 2.9%, 2.2%, 3.8%, and 5.0% in the four data sets. The proposed method significantly improves object detection accuracy for organisms in complex marine environments.