Key Lab of Industrial Computer Control Engineering of Heibei Province, Yanshan University, Qinhuangdao 066004, China.
Key Lab of Industrial Computer Control Engineering of Heibei Province, Yanshan University, Qinhuangdao 066004, China.
Mar Pollut Bull. 2024 Aug;205:116644. doi: 10.1016/j.marpolbul.2024.116644. Epub 2024 Jul 2.
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.
海洋垃圾清理是海洋环境保护的一个紧迫问题。具有视觉识别技术的 AUV 逐渐成为一个中心研究问题。然而，现有的识别算法推理速度慢，计算开销大。它们还受到模糊图像和干扰信息的影响。为了解决这些问题，提出了一个实时语义分割网络，称为 WaterBiSeg-Net。首先，我们提出了多尺度信息增强模块来解决低清晰度和模糊图像的影响。然后，为了抑制背景信息的干扰，提出了门控聚合层。此外，我们还提出了一种可以直接提取边界信息的方法。最后，在 SUIM 和 TrashCan 数据集上进行了广泛的实验，结果表明 WaterBiSeg-Net 可以更好地完成海洋垃圾分割任务，并为 AUV 实时提供准确的分割结果。这项研究为 AUV 识别海洋垃圾提供了一种低计算成本和实时的解决方案。
