Single Image Rain Removal Based On Squeeze-and-Excitation Networks

With the development of technology and the economy,computer vision systems are widely applied to a lot of fields.Such as traffic surveillance system,security monitoring system,industrial supervision system and so on.However,bad weather can affect the outdoor computer vision systems easily.Under bad weather condition,images or videos captured by outdoor computer vision systems might suffer from severe degradation,which could lead to the failures of some computer vision algorithms.As the most common bad weather,rain has a particularly wide impact on image quality,and it leads to the rise of studies of rain removal.Recently,most studies focus on rain removal from videos,which is different from single image rain removal since videos contain lots of temporal information while single images have none.Hence,it is more challenging to remove rain from single image.This paper focuses on single image rain removal,and the main works and innovations are listed as follows:(1)This paper proposes a novel method for neural network connection,which called reusing original information connection.By reusing the original input,it can provide more background details for the network.And the reconstruction process after rain streak removal can benefit from these details.Experimental results show the effectiveness of this kind of connection.(2)In this paper,we analyze the ways that some state-of-the-art methods used for decomposing detail layers from images.We train one of our proposed networks with two kinds of inputs,one is the rainy images directly,and the other is the detail layers.Then we analyze the differences of the rain removal results between these two different training inputs.To our knowledge,this is the first deep learning based rain removal work to conduct an explicit comparison between these two kinds of inputs.(3)Based on squeeze-and-excitation block,we propose two novel networks to address the problem of single image rain removal.The first one is a multi-scale feature learning network,the other one is a squeeze-and-excitation network with reusing original information connection.Most of the deep learning-based rain removal methods focus mainly on the spatial information while ignoring the channel-wise relationship.The networks proposed in this paper take both spatial and channel-wise information into account via squeeze-andexcitation block,which is designed to explicitly model the interdependencies between the channels of feature maps.Experimental result show that the proposed networks perform well even with very shallow layers.