基于残差注意力机制的泥石流沟谷识别  

Debris flow gully recognition based on residual attention mechanism

作　　者：刘坤香 王保云 徐繁树 韩俊 LIU Kunxiang;WANG Baoyun;XU Fanshu;HAN Jun(School of Information,Yunnan Normal University,Kunming,Yunnan 650500,China;School of Mathematics,Yunnan Normal University,Kunming,Yunnan 650500,China;Key Laboratory of Complex System Modeling and Application for Universities in Yunnan,Kunming,Yunnan 650500,China)

机构地区：[1]云南师范大学信息学院,云南昆明650500 [2]云南师范大学数学学院,云南昆明650500 [3]云南省高校复杂系统建模及应用重点实验室,云南昆明650500

出　　处：《中国地质灾害与防治学报》2022年第6期134-141,共8页The Chinese Journal of Geological Hazard and Control

基　　金：国家自然科学基金项目(61966040)。

摘　　要：针对泥石流灾害沟谷图像分类问题,文章对Resnet18网络进行改进,提出了一种改进的卷积神经网络模型。通过在网络结构中加入残差注意力模块,解决了原模型提取图像特征较差、边缘模糊的问题,改进后的网络能精确捕捉到泥石流灾害沟谷图像中的轮廓和内部山脊信息。此外,文章还对多种注意力机制结构进行了实验对比,分析其差异性,得出最适合泥石流灾害沟谷数据分类的注意力机制网络。实验表明改进后的网络模型在泥石流灾害沟谷图像的分类准确率达到75.42%,其分类性能在Resnet18网络模型的基础上提升了5.1%。For debris flow disasters in valleys image classification problems,this paper improved the Resnet18 network,an improved convolution neural network model is put forward,through adding residual attention in network structure module,solved the original model to extract the image features to solve the problem of poor,edge model accurately capture the debris flow disasters in valleys in the image contour and internal ridge information.In addition,this paper also conducts comparative experiments on various attention mechanism structures,analyzes their differences,and obtains the attention mechanism network most suitable for debris flow disaster gully data.The experimental results show that the classification accuracy of the improved network model in debris flow disaster gullies reaches 75.42%,and its classification performance is improved by 5.1%compared with the Resnet18 network model.

关 键 词：Resnet18 注意力机制 遥感影像 泥石流灾害 

分 类 号：P642.23[天文地球—工程地质学]