Research On Natural Image Shadow Removal Based On Generative Adversarial Networks

In natural images,shadows are almost everywhere.Shadows can provide us with information such as object position and lighting conditions,but also bring difficulties to many image processing tasks.In order to reduce the influence of shadows on image processing and make subsequent tasks proceed smoothly,it is very important to study a fast and effective shadow removal algorithm.The calculation process of early shadow removal research is complicated,and the effect of shadow removal is not ideal.With the rapid development of deep learning,especially the high performance and high efficiency when processing massive data such as images,traditional shadow removal algorithms are gradually replaced by deep learning methods.The advent of generative adversarial networks has injected new vitality into computer vision.Nevertheless,the existing shadow removal algorithms still have problems such as obvious shadow boundaries and poor restoration of image details,etc.,which makes the shadow removal results not improved.In addition,due to the difficulty of labeling shadow masks,the number of currently available triplet shadow datasets is not large enough to train and evaluate deep network models well.In response to the above problems,this paper mainly made the following three contributions.Firstly,in order to train and evaluate the model better,it is necessary to enrich and supplement the existing shadow dataset.To this end,we shot a new set of shadow data sets,and used Photoshop to mark the shadow mask.Finally,a new data set of triples named TIDSR is sorted out.Secondly,in order to make the image processing more refined and to further improve the effect of shadow removal,this paper proposes SRNet,a generational adversarial shadow removal model that combines multi-scale discrimination and spectral normalization discrimination.The multi-scale discriminator trains on images of different resolutions,perceives image features of different granularities,and thus restores the image on different levels of detail.At the same time,spectral normalization is added to the model to limit the discriminator parameters to solve the problem of unstable network training.Finally,in order to solve the problem of training data mismatch and reduce the use of prior information,this paper further builds a shadow removal model based on CycleGAN.The original CycleGAN pays more attention to the macroscopic features of the image and therefore does not handle fine-grained details.Through a series of experiments and comparisons,we finally propose an unsupervised CycNet model that does not require paired data.CycNet also adds multi-scale discrimination to extract and restore the detailed features of different levels of the image to improve the shadow removal effect.In this paper,the results of the proposed model and other existing shadow removal models are qualitatively compared and quantitatively analyzed.The experimental results show that the shadow removal effect of SRNet has been further improved.Multi-scale discriminator plays an extremely important role in improving the effect of shadow removal.CycNet achieved similar results to SRNet in independent training.Although the effect is slightly reduced in cross-training,it greatly reduces the requirements for training data and provides a new idea for image shadow removal research.