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

Area of Science:

• Marine environmental protection
• Robotics
• Computer Vision

Background:

• Marine debris poses a significant threat to ocean ecosystems.
• Autonomous Underwater Vehicles (AUVs) equipped with visual recognition are crucial for marine debris cleanup.
• Current recognition algorithms suffer from slow speeds, high computational costs, and are susceptible to image quality issues.

Purpose of the Study:

• To develop a real-time semantic segmentation network for accurate marine debris identification by AUVs.
• To address limitations of existing algorithms, including slow inference and susceptibility to image degradation.
• To provide a computationally efficient solution for marine debris segmentation.

Main Methods:

• Proposed WaterBiSeg-Net, a real-time semantic segmentation network.
• Introduced the Multi-scale Information Enhancement Module to mitigate effects of low-definition and blurred images.
• Developed the Gated Aggregation Layer to reduce background interference.
• Implemented a novel method for direct extraction of boundary information.

Main Results:

• WaterBiSeg-Net demonstrated superior performance in marine debris segmentation on SUIM and TrashCan datasets.
• The network provides accurate segmentation results essential for real-time AUV operations.
• Achieved efficient and accurate identification of marine debris in challenging underwater conditions.

Conclusions:

• WaterBiSeg-Net offers a low computational cost and real-time solution for AUVs to identify marine debris.
• The proposed network effectively handles blurred images and background noise, improving segmentation accuracy.
• This research advances the capabilities of AUVs in marine environmental protection efforts.