Abstract:The cleanup of marine debris is an urgent problem in marine environmental protection. AUVs with visual recognition technology have gradually become a central research issue. However, existing recognition algorithms have slow inference speeds and high computational overhead. They are also affected by blurred images and interference information. To solve these problems, a real-time semantic segmentation network is proposed, called WaterBiSeg-Net. First, we propose the Multi-scale Information Enhancement Module to solve the impact of low-definition and blurred images. Then, to suppress the interference of background information, the Gated Aggregation Layer is proposed. In addition, we propose a method that can extract boundary information directly. Finally, extensive experiments on SUIM and TrashCan datasets show that WaterBiSeg-Net can better complete the task of marine debris segmentation and provide accurate segmentation results for AUVs in real-time. This research offers a low computational cost and real-time solution for AUVs to identify marine debris.

WaterBiSeg-Net: An underwater bilateral segmentation network for marine debris segmentation