Research On Image Shadow Removal Based On Generative Adversarial Networks

In this paper,a progressive attention mechanism guided image shadow removal algorithm is proposed to solve the problems of residual shadow and incomplete shadow removal of complex objects or dark areas in the process of image shadow removal.The algorithm is studied and designed in the framework of generative adversarial network.Firstly,in the feature extraction stage of generating network,we use hole convolution residual blocks with different learning rates to extract features,expand the receptive field of the network,and then improve the accuracy of external attention mechanism in extracting shadow location and contour information;Then,the parallel attention mechanism is used in the self encoder of the generation network to guide the self encoder to encode the detail information of the non shadow area in the shadow image.Secondly,in the self coding stage,multi-layer and multi-scale feature fusion method is used to make the coding take into account the global semantic information and local detail features;finally,the series attention mechanism is used to guide the discrimination network to identify the shadowless image generated by the generated network,so as to reduce the loss of key features and enhance the identification ability of the discrimination network.The algorithm is tested on the open data set SRD and ISTD,and the visual effect is good.The SSIM value of the algorithm can reach 97.5%,the PSNR can reach 32.8,and RMSE can reach 6.36.Compared with similar algorithms,it has certain advantages.After shadow removal,the image feature information is preserved completely,the image picture is clear,and the shadow removal effect of complex ground objects or green plants and dark areas is obvious.The paper has 36 pictures,7 tables,and 52 references.