Research On Multi-scale Based Single Image Deraining

With the development of the times,as one of the key fields of artificial intelligence,computer vision is at the core of the fourth scientific and technological revolution,and has attracted widespread attention.Nowadays,various computer vision algorithms,such as,object recognition,object detection,semantic segmentation,target motion tracking,and visual question answering,etc.,have been explored and achieve promising results.These computer vision algorithms are widely used in the military and transportation fields.However,as one of the common bad weather condition,the quality of the outdoor photos or videos collected on rainy days is often impaired due to the occlusion of rain streaks,which degrade the performance of the subsequent computer vision algorithms.Therefore,removing rain streaks from video or image is a necessary but challenging pretreatment process,and it has a very wide range of applications in the real life.According to the different inputs,the problem of rain removal is divided into video deraining and single image deraining.This thesis focuses on the latter.The traditional single image deraining algorithms are generally based on the image decomposition model.This type of algorithm ignores the low frequency components of the image.Thus it is either difficult to distinguish between rain atoms and non-rain atoms with dictionary training,or causing blur in the deraining output.With the development of deep learning,neural networks are employed in various application fields of image processing.Recently,a large number of deep learning-based algorithms remove rain streaks by constructing different network models,and studying the non-linear mapping from rainy pictures to clean pictures.Compared with the traditional methods,the performance of deep learning based methods is better.However,since the rain streaks in the image usually are different in sizes and directions,and overlap with each other.One of the problems of these methods is that they cannot often distinguish heavy rain and light rain,and it is easy to ignore and fail to restore the texture information in the image.As a result,it further lead to problems such as excessive/insufficient rain removal,and the missing of the background information.Different from the existing methods which remove rain streaks directly in the original image resolution.In this thesis,we propose a single image deraining methodbased on the residual multi-scale model.Firstly,we down-sample the input rain image to different resolutions.This operation will make the heavy rain in the original resolution smaller in the low-resolution image,and the light rain will be removed automatically.Secondly,we remove the rain gradually from the rain image from lowest to highest resolutions.In particular,we up-sample the low-resolution deraining result with the factor of 2?,and calculate the residual map between it and the rain image with the higher resolution.Then we concat the residual map with the rain image and input them into the model in the next stage.Finally,this process is repeated according to the factor of down-sampling,and we get the final deraining result with the original resolution.In this process,the deraining models in the different stages are trained independently.As the further improvement of the proposed method,we employ recurrent mechanism and an adaptive receptive selection mechanism to lighten the model.The improved deraining network could adjust its recepetive field in the different deraining stages.It not only improves the training efficiency,but also improves the performance of the deraining.A large number of experiments are carried out in the synthetic datasets Rain12000,Rain14000,Rain H,Rain L,Rain800 and real-world dataset Real15.In the synthetic datasets,the experimental results show that the proposed method outperforms state-of-the-art deraining methods.In the real-world datasets,the rain-free images generated by the proposed method have achieved better visual effects.