Research And Implementation Of Image Deraining Method For Real Scenes

Currently,the field of image deraining lacks pairs of rain/clean images in real scenes.Deep learning-based methods,usually trained using largescale pairwise synthetic datasets,have achieved remarkable progress.However,current synthetic datasets cannot simulate real rain streaks very well.Applying deraining models trained on synthetic datasets directly to real rain images often does not perform well.In addition,the existing deraining models still have problems such as blurred background and residual rain streaks in the restored image after deraining.Therefore,the problem of deraining images in real scenes is still a challenge.In order to solve this problem,this paper proposes an image deraining method for real scenes,which can extract and utilize rain features from synthetic rain images and real rain images,improve the deraining effect on real rain images,and restore a clear,rain-free background for real rain images.More specifically,firstly,this paper synthesizes design ideas from following two aspects,and proposes an image deraining model for real scenes.On the one hand,this paper designs a semi-supervised image deraining model,which is divided into two parts:supervised learning part and unsupervised learning part.The supervised learning part uses paired synthetic datasets for training,and the unsupervised learning part uses unpaired real datasets for training.The two parts are mutually constrained to obtain the final training and learning results.At the same time,in the unsupervised learning part,a multi-layer,image-level contrast loss is introduced,and the real rain image and the unpaired clear image without rain are used as negative samples and positive samples,respectively,to ensure that the deraining real rain image is close to the unpaired clear image without rain in the feature space,and far away from the real rain image.On the other hand,this paper designs a multi-stage image deraining model based on attention mechanism for modeling the existing image blurring and rain streak residual problems in current deraining models.The network architecture of the model is based on encoder-decoder.The encoder is designed based on Swin Transformer to extract the global features of image.A self-supervised memory module is introduced between the encoder and decoder to store rain information.At the same time,the model introduces a multi-stage mechanism.In this paper,two stages are used in the implementation.The rough features are extracted in the first stage,and the features are integrated and improved in the second stage.Using the information interaction between the two stages,the rain image can restore clear background without rain step by step.In summary,the semisupervised image deraining model and the multi-stage image deraining model based on the attention mechanism together constitute an image deraining model for real scenes.The semi-supervised image deraining model is a training method of the image deraining model for real scenes.The network of the multi-stage image deraining model constitutes the network of the supervised learning and unsupervised learning parts of the semi-supervised image deraining model.In addition,the real rain image datasets used in current papers often have a relatively small amount of data,which is not enough to cover various complex situations in the real world.This paper summarizes the real datasets used in current papers,and uses Google search to construct a real rain image dataset containing 1100 unlabeled real rain images named RealRain11k.Next,the existing evaluation indicators cannot accurately evaluate the quality of the restored image after the real rain image has been derained.In addition to using the existing evaluation methods,this paper uses YOLOv5 target detection algorithm to evaluate the effect of deraining model based on its accuracy before and after deraining the real rain image.Extensive experiments and research on several datasets(such as Rainl3k,Rain800,ReakRainllk)show that compared with existing deraining models,the proposed image deraining model for real scenes has better performance on real datasets.Finally,on the basis of above research,an image/video rain removal system is designed and implemented,which is convenient for users to use.