Research Of Single Image Rain Streak Removal Based On Multi-scale Joint Feature

In daily life,people are often affected by bad weather,such as fog,snow and rainstorm.These severe weather not only affect people's normal travel,but also causes interference and degradation of images or video.The research objective of this paper is how to reduce the impact of these severe weather on the computer vision system and restore the image interfered by the severe weather.Due to the development of deep convolutional neural network,the research of single image de-raining has made great progress recently.However,the existing methods all regard convolutional neural network as an encapsulated end-to-end mapping module,and do not deeply discuss the rationality and superiority of neural network design.In this paper,an effective end-to-end neural network structure is studied to enhance the expression of features and the characteristic connection between the front and back network Spaces.Particularly,we propose a Encoder-Decoder network framework(FJDN)for extract multiscale aggregation features.The whole framework of the network is composed of a EncoderDecoder network,and each Encoder block corresponds to a Decoder block.The Encoder network conducts down-sampling through the maximum pooling layer,and records the maximum pooled index position in the process of maximum pooling.The pooling indices will guide the up-sampling recovery process of the corresponding Decoder network.Specifically,the Encoder network and the Decoder network are same in the convolutional layer,both of which are Feature Joint Dense Block(FJDB).Only the maximum pooling and its corresponding up-sampling process are different.In this way,the process of down-sampling and up-sampling can effectively learn more abstract features,so as to perfectly depict the features of the rain streaks and better restore the background detail.The rain removal algorithm proposed in this paper has done a lot of experiments on the synthetic data set and the real data set,and the experiment has proved that the method can effectively remove the rain streaks,while well restoring the details of the image.In addition,the proposed algorithm is also applied to the image dehazing problem,and excellent results are obtained.