基于条件生成对抗网络的输电线路巡检图像水痕去除方法

打开文本图片集

中图分类号：TP391 文献标志码：A

Abstract：Toachieve high-quality water stainremovalfrom transmission line inspection imagesandefectively mitigate theadverseimpactof waterstainocclusion，amethodofwaterstainremovalfortransmision lineinspection images based onconditional generativeadversarial network was proposed.Theoriginal water-stained imageswereusedas inputs，anda generator incorporating an atention mechanism was constructed based on U-Net architecture to enhance the generator’s ability to focus on critical water stain features.Meanwhile，adiscriminator was employed to supervise the authenticityof thegeneratedwater-stain-free images，therebyimproving training stabilityandqualityofthegeneratedwater-stain-free images.A totalof985pairsoforiginal water-stained images andcoresponding groundtruth water-stain-free images were partitioned in a 4：1 ratio of a training set to a test set based on the number of image pairs，followed by water stain removal validationandobjectdetection experiments.Theresultsshowthatthe proposedwaterstainremoval method significantly improvesthequalityof thegenerated water-stain-freeimages，achievinga structural similarityindex measureof 0.867 compared to the ground truth water-stain-free images，andthe generated water-stain-free images moreaccurately match theground truth water-stain-fre images compared totheimages generatedby pix2pix network andcycle-consistent generativeadversarial network．Whenapplied totheobjectdetectionalgorithmfor identifying bird nestsand suspendedobjects， theproposed method achievesan average precision increaseof17.2and19.1 percentage points，respectively，compared tousing theoriginal water-stained images，significantly improving thedetection performanceforlow-quality transmission line inspection images obscured by water stains.

Keywords：image water stainremoval method；conditional generativeadversarial network；attention mechanism；transmission line inspection image；object detection

输电线路巡检图像（后文中在不至于引起歧义的情况下简称巡检图像）是无人机巡检中最有价值的信息载体之一。（剩余11439字）