Underwater instance segmentation: a method based on channel spatial cross-cooperative attention mechanism and feature prior fusion

Abstract

In aquaculture, underwater instance segmentation methods offer precise individual identification and counting capabilities. However, due to the inherent unique optical characteristics and high noise in underwater imagery, existing underwater instance segmentation models struggle to accurately capture the global and local feature information of objects, leading to generally lower detection accuracy in underwater instance segmentation models. To address this issue, this study proposes a novel Channel Space Coordinates Attention (CSCA) attention module and a Channel A Prior Attention Fusion (CAPAF) feature fusion module, aiming to improve the accuracy of underwater instance segmentation. The CSCA module effectively captures local and global information by combining channel and spatial attention weight, while the CAPAF module optimizes feature fusion by removing redundant information through learnable parameters. Experimental results demonstrate significant improvements when these two modules are applied to the YOLOv8 model, with the mAP@0.5 metric increasing by 3.2% and 2% on the UIIS underwater instance segmentation dataset. Furthermore, the instance segmentation accuracy is significantly improved on the UIIS and USIS10K datasets after these two modules are applied to other networks.