融合残差注意力机制的深度可分离UNet泥石流堆积扇分割  

Segmentation of debris flow fan by a depth-separable UNet incorporating a residual attention mechanism

作　　者：宋昕 王保云 朱双颖[3] SONG Xin;WANG Baoyun;ZHU Shuangying(School of Mathematics,Yunnan Normal University,Kunming 650500,China;Yunnan Key Laboratory of Modern Analytical Mathematics and Applications,Yunnan Normal University,Kunming 650500,China;School of Architectural Engineering,Kunming University,Kunming 650500,China)

机构地区：[1]云南师范大学数学学院,云南昆明650500 [2]云南师范大学云南省现代分析数学及其应用重点实验室,云南昆明650500 [3]昆明学院建筑工程学院,云南昆明650500

出　　处：《辽宁工程技术大学学报(自然科学版)》2024年第6期112-120,共9页Journal of Liaoning Technical University (Natural Science)

基　　金：国家自然科学基金项目(61966040);云南省科技厅地方高校联合专项(202001BA070001-171)

摘　　要：为解决传统机器学习方法在泥石流堆积扇识别中的精度低、效果差的问题,提出一种基于深度学习的残差注意力可分离UNet算法(RAMS-UNet)。该算法在编码部分采用VGG16主干网络进行特征提取,加深网络层次;在跳跃连接部分引入改进的注意力机制,强化信息传递;在解码部分使用深度可分离卷积和密集连接块,进一步增强空间和通道上的信息表达能力。研究结果表明:与其他算法相比,RAMS-UNet算法对泥石流堆积扇的分割精度更高,mIoU、mPA、PA和F1指数等评价指标均显著提升。RAMS-UNet算法突破了传统方法在泥石流堆积扇识别中的局限性,为泥石流灾害评估提供了更加精准的信息支持。In order to solve the problem of low accuracy and poor effect of traditional machine learning methods in debris flow accumulation fan recognition,a residual attention separable UNet algorithm based on deep learning(RAMS-UNet)is proposed.The algorithm uses VGG16 backbone network for feature extraction in the coding part to deepen the network level;an improved attention mechanism is introduced in the jump connection part to strengthen information transmission.In the decoding part,deep separable convolution and dense connection blocks are used to further enhance the information expression ability on space and channels.The results show that compared with other algorithms,the RAMS-UNet algorithm has higher segmentation accuracy for debris flow accumulation fans,and the evaluation indexes such as mIoU,mPA,PA and F1 index are significantly improved.The RAMS-UNet algorithm breaks through the limitations of traditional methods in the identification of debris flow fans,and provides more accurate information support for debris flow disaster assessment.

关 键 词：泥石流堆积扇 沟谷型泥石流 语义分割 UNet算法 注意力机制 深度可分离卷积 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]