基于改进ResNet50网络的泥石流沟谷识别  

Debris Flow Gully Identification Based on Improved ResNet50 Network

作　　者：刘秋雨 王保云 LIU Qiuyu;WANG Baoyun(School of Mathematics,Yunnan Normal University,Kunming,Yunnan,China 650500;Yunnan Key Laboratory of Modern Analytical Mathematics and Applications,Yunnan Normal University,Kunming,Yunnan,China 650500)

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应用重点实验室,云南昆明650500

出　　处：《昆明学院学报》2024年第6期114-119,128,共7页Journal of Kunming University

基　　金：国家自然科学基金项目(61966040).

摘　　要：针对传统神经网络进行泥石流沟谷图像分类时,可能出现准确率不高、提取图像特征较差、边缘模糊等问题,对ResNet50网络进行改进.在ResNet50网络部分残差块前加入注意力机制模块,使其具有更高的性能和准确性,可以精确捕捉到泥石流沟谷图像中的地形地貌.试验结果表明,改进后的ResNet50网络在泥石流沟谷图像的分类准确率达到83.02%,其分类性能在ResNet50网络的基础上提升了11.32个百分点,且准确率、召回率、精确率、F 1值和AUC值等各项指标均优于ResNet50网络和其他深度学习识别算法.Traditional convolutional neural networks have problems such as low accuracy,poor image feature extraction and blurred edges when used for landslide disaster valley image classification.This paper improves the ResNet50 network by adding a CBAM attention mechanism module before some residual blocks of the ResNet50 network,which enables it to have higher performance and accuracy and accurately capture the terrain and landforms in landslide disaster valley images.The experimental results show that the improved ResNet50 network achieves a classification accuracy of 83.02%for landslide disaster valley images,which improves its classification performance by 11.32 percentage points compared to the ResNet50 network.Moreover,its accuracy,recall rate,precision rate,F1 value,and AUC value are better than those of the ResNet50 network and other deep learning recognition algorithms.

关 键 词：注意力机制 卷积神经网络 泥石流灾害 ResNet50 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术] X928.7[环境科学与工程—安全科学]