基于改进CycleGAN的轨道扣件ISR方法  

Image Shadow Removal Method for Rail Fastener Based on Improved CycleGAN

作　　者：孙践知[1] 吴浩[1] 杨亚峰 于重重[1] SUN Jian-zhi;WU Hao;YANG Ya-feng;YU Chong-chong(College of Computer Science,Beijing Technology and Business University,Beijing 100048,China)

机构地区：[1]北京工商大学计算机学院,北京100048

出　　处：《计算机仿真》2023年第5期208-212,共5页Computer Simulation

基　　金：京沪高铁基础设施无人机巡检关键技术研究(铁路总公司,I20D00010)。

摘　　要：围绕国家对高铁基础设施安全保障的重大需要,基于无人机图像的高铁基础设施缺陷检测可以作为目前巡检手段的补充和替代。轨道扣件是高铁基础设施巡检中重要的工务巡检对象,但无人机拍摄图像中存在高铁轨道两侧电气塔杆的阴影投射遮挡轨道扣件的情况,严重影响检测效率和精确度。针对以上问题,提出一种基于改进CycleGAN的轨道扣件图像去阴影方法,将掩码机制加入到CycleGAN中提高无监督学习方法对阴影区域的关注,同时将谱范数归一化加入到网络中以稳定训练过程、提升去阴影效果。在公共阴影数据集USR和无人机拍摄高铁扣件数据集上的实验结果表明,改进CycleGAN较好地提升了无人机图像去阴影效果。In view of the country's great need for the security guarantee of high-speed railway infrastructure,the defect detection of high-speed railway infrastructure based on UAV image can be used as a supplement and substitute for the current inspection means.Track fastener is an important inspection object in high-speed railway infrastructure inspection.However,the shadow projection of electrical tower poles on both sides of high-speed railway tracks in the images taken by UAV blocks track fastener,which seriously affects the detection efficiency and accuracy.To solve the above problems,this paper proposes an improved CycleGAN image shadow removal method for track fasteners.In this method,the mask mechanism is added to CycleGAN to improve the unsupervised learning method's attention to the shadow region,and the spectral normalization is added to the network to stabilize the training process and improve the shadow removal effect.Experimental results on USR and UAV-shot high-speed rail fasteners data sets show that improved CycleGAN can better improve uav image shadow removal.

关 键 词：轨道扣件图像 图像去阴影 掩码机制 谱范数归一化 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]