DSDMNet: A Dual-Supervised Distraction Mining Network for Underwater Image Segmentation

Underwater Image Segmentation (UIS) usually encounters challenges such as turbidity, light scattering and underwater biological camouflage. Current UIS methods often focus on tackling the turbidity and light scattering conditions while ignoring the influence of biological camouflage. To address the complexities inherent in UIS tasks, we propose a Dual-Supervised Distraction Mining Network (DSDMNet). Specifically, we utilize PVTv2 as the backbone network to conduct preliminary feature extraction. Subsequently, we introduce a Cross-Scale Atrous Pyramid Aggregation Module (CSAPAM) to construct feature pyramids using atrous convolution based on the extracted features. Finally, we develop a Dual-Supervised Distraction Decoder (DSDD) that refines the segmentation results by fusing the background and foreground features, guided by the background and foreground mask supervisions, respectively. Comprehensive experimental evaluations on the USOD10K data set reveal that DSDMNet consistently outperforms nine baseline methods.