Shadow Removal Of Single Image Based On Cycle Generative Adversarial Network

The shadow is a physical phenomenon that represents the extent of darkness that occurs when light is blocked by an opaque object.Shadow has a lot of negative effects on production and daily life.For example,when image collection is carried out in the fields of transportation,terrain detection,security and defense,shadow will seriously interfere with the collected image,resulting in information loss,color change,insufficient brightness and other problems in the image.Therefore,eliminating the shadow in the image is of great significance for improving the accuracy of image recognition and image segmentation.To solve the above problems,this paper focuses on the shadow removal method of a single image.The specific work includes:1.Aiming at the problem that the traditional deep learning image shadow removal method training is limited by the matching of the standard data,a new shadow removal method based on the gradient constraint is proposed.Firstly,a new unpaired data set is constructed to train the model.Secondly,the generator and discriminator of the whole structure are constructed by selecting the generator and discriminator in Wasserstein generated countermeasure network(WGAN).Finally,the loss function of the generator is optimized through WGAN Gradient Penalty[30],so as to achieve the goal of removing shadows.Compared with the existing methods of deep learning to remove shadows,this method does not need to use paired data sets for training,which provides a new idea for the preparation of data sets.Experiments show that this method is better than the methods of Guo et al.[10]and Gong et al.[8]in removing the shadow of a single image,with the root-mean-square error value of 8.03.2.In view of the image shadow removal method based on traditional cycle generative adversarial network is easy to produce unnatural traces when eliminating the shaded part[27],a shadow removal method based on spectral normalized cycle generative adversarial network is proposed.The network consists of three parts:the first part,generating non-shadow image through inputting the real shadow image;In the second part,generating the shadow mask through inputting the shadow image;The third part,generating shadow image through inputting the real non-shadow image.Then,the loss function of generators is optimized through Spectral normalization[31].Finally,a new unpaired data set is used to train the network model to perform shadow removal on a single image.Experiments show that the method proposed in this paper has a root-mean-square error value of 7.86,which is better than the traditional cycle generative adversarial network.In addition,it solves the problem that the traditional cycle generative adversarial network shadow removal method doesn’t has a good effect on the preservation of image background material information,our method has a good effect on the preservation of texture information.3.According to the above research content,a single image shadow removal system based on cycleGAN is designed and implemented.The system integrates the shadow removal method proposed in this paper.