UW-DETR: Underwater Image Instance Segmentation with Co-DETR

Underwater image instance segmentation is crucial for marine environmental protection and management. However, the light scattering effect of the underwater environment can reduce the image quality and blur the boundaries of objects, which can significantly impair the performance of underwater instance segmentation. Furthermore, there are organisms that are camouflaged underwater and can perfectly blend into the underwater environment. To address these issues, we propose UW-DETR, a Co-DETR-based instance segmentation model for underwater images. Firstly, the Boundary Aware Module (BAM) to generate a priori edge information, thereby providing effective edge information in underwater images. Subsequently, the a priori edge information is combined with the feature maps, which serves to enhance the model's boundary perception, reduce the impact of scattering effects on the image, and improve the detection of camouflaged targets, thereby aiding the model's prediction. Secondly, we were inspired by the Boundary Learning Loss (BLL) in WaterMask, which highlighted the importance of boundary loss in underwater instance segmentation tasks. However, BLL does not consider the boundary loss of underwater camouflaged targets, so we improved on BLL and proposed Multiple Boundary Losses (MBL) to guide the model to produce more accurate results. A series of experiments conducted on the UIIS dataset demonstrated that our method achieved a mean average precision (mAP) of 29.2 using ResNet-101 as the backbone, representing an improvement of 2AP over the WaterMask R-CNN.