基于改进深度学习的航拍滑坡检测方法  

Aerial landslide detection method based on improved deep learning

作　　者：杨靛青 毛艳萍 YANG Dian-qing;MAO Yan-ping(School of Economics and Management,Fuzhou University,Fuzhou 350116,China)

机构地区：[1]福州大学经济与管理学院,福建福州350116

出　　处：《计算机工程与设计》2024年第1期268-274,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(71572040);福建省自然科学基金面上基金项目(2021J01573);福建省社会科学规划基金项目(FJ2019B139)。

摘　　要：为及时发现滑坡险情展开应急救援,提出一种结合逐通道不同阈值的深度残差收缩网络(DRSN-CW)方法的更快速区域卷积神经网络(Faster-RCNN)模型的航拍图像滑坡检测算法。利用图像增强的伽马变换、高斯滤波方法提高图片的质量;使用群组归一化方法消除batchsize大小对模型的影响;为减少噪声以及无关特征区域的干扰,采用DRSN-CW网络对滑坡检测目标进行精准定位。实验结果表明,改进后模型与先前模型相比F1值以及平均精度分别增加了10.7%、10.2%,可以有效检测滑坡险情。To detect landslide danger in time and carry out emergency rescue,an UAV aerial image landslide detection algorithm based on the Faster-RCNN model of the deep residual shrinkage networks with channel-wise thresholds(DRSN-CW)method was proposed.The image enhancement Gamma transform and Gaussian filtering method were used to improve the quality of the picture.The group normalization method was used to eliminate the effect of batchsize on the model.To reduce the noise and interference from irrelevant feature areas,the DRSN-CW network was used to precisely locate the landslide detection target.Experimental results show that the F1-score and average accuracy of this improved model are increased by 10.7%and 10.2%compared with the previous model.

关 键 词：更快速区域卷积神经网络 滑坡检测 无人机航拍 伽马变换 高斯滤波 群组归一化方法 逐通道不同阈值的深度残差收缩网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]