改进U-Net的水岸分割算法  

Improve U-Net for Watershore Segmentation

作　　者：姚福飞 柏利志 周田[1] 

机构地区：[1]同济大学,上海

出　　处：《计算机科学与应用》2022年第12期2875-2883,共9页Computer Science and Application

摘　　要：我国水域广阔、多样等特点导致难以使用人工的方式实现全面高效地监测、管理,使用现代信息技术辅助水域管理可以极大地提高效率。水岸分割是水域管理信息化的关键技术,传统的水岸分割算法依赖手工特征提取需要针对特定的环境调整参数,适应性不高不适用于多变的天气条件和复杂的河岸场景。基于学习的方法比传统方法更能适应不同的环境,本文使用基于深度学习的U-Net模型进行水岸分割,并引入空间注意力模块调整不同空间的特征权重,使用图像处理方法模拟环境中的雨雾进行数据增强。实验结果表明模型分割MIoU达97.60%,较已有的算法性能有显著提升。所提模型在极端挑战环境中仍能保持较高的分割准确率,与传统算法相比具有优秀的环境适应性。The characteristics of the vast and diverse waters in the country make it difficult to realize efficient monitoring and management manually. Modern information technology has potential for improving the efficiency of water management greatly. Watershore segmentation is the key technology for water management informatization. Traditional watershore segmentation algorithms rely on manual feature extraction and need to adjust parameters for specific environments, which can’t be used in watershore segmentation in a variety of weather conditions and in complex inland river scenarios due to its poor environment adaptability. Compared with traditional methods, learning-based methods are more adaptable to different environments. In this paper, U-Net is used for the water bank segmentation, which is a deep learning method. The spatial attention module is introduced to adjust the feature weights of different spaces, and the image processing method is used to simulate the rain and fog in the environment for data enhancement. The experimental results show that the model segmentation MIoU is 97.60%, which is significantly improved compared with the existing algorithm. The proposed model maintains high accuracy even in extremely challenging environments, having the property of impressive environmental adaptability compared with the traditional algorithm.

关 键 词：水域管理 水岸分割 U-Net 注意力 适应性 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]