Deep Models For Single Image Deraining

Clean images can provide guaranteed input signals for video surveillance,target tracking,target detection,and autopilot,ensuring that these tasks can function normally.However,rain is a common weather in daily life,and images taken on rainy days severely reduce visual effects,which can also cause those visual tasks to fail.Therefore,it is a problem to be solved to recover a clear image from a rainy image.The shape,size and direction of the rain line in a rainy day also lead to a very challenging problem to remove rain from the image.Therefore,an effective rain removal algorithm should be proposed.This paper solves the problem of removing rain from a single image from multiple angles.1)Considering the importance of spatial context information to rain removal,we propose a more effective deraining unit to solve the problem of rain removal in a single image.In this deraining unit,we use the dilation convolution and the Squeeze-and-Excitation operations to obtain more spatial context information and semantic relevance between channels,respectively.Based on the dilation convolution with different factors,we can obtain multi features at different levels.Based on the Squeeze-and-Excitation operation,each channel can be adaptively assigned a weight to obtain the semantic correlation between channels.Finally,we connect multiple deraining units in a dense manner to maximize the flow of feature information at different levels.Our designed deraining unit and dense connections enable our network to have strong rain streaks representation ability.2)In the real-world rainy image,the existing rain removal algorithms usually leave a large number of rain streaks,which is difficult to receive for applications.Our method is inspired by a natural idea that a good deraining method should have the ability to process various rain streaks in a repeated manner until the rain streaks are removed out cleanly.This also implies that estimated deraining image,in which there are remaining rain streaks,can be further processed by the deraining method.So our proposed deraining approach is in a coarse-to-fine manner that is multiple stages.What's more,the model size keeps in a small weight due to that the parameters are shared in each stage.To deal with kinds of rain streaks,we propose a densely connected dilation convolution block.The dilation convolution can enlarge the receptive field,which can acquire more spatial contextual information,to handle with rain streaks with different sizes.In the block,a densely connected style is adopted,which can maintain important features from different levels.Moreover,we find that the fusion among all the deraining results in previous each stage can promote the subsequent deraining performance.We think that the deraining results in previous stages can guide the subsequent deraining procedures.3)As spatial contextual information is important for single image de-raining,we develop a multi-scale kernels de-raining layer,which can utilize the multi-scale kernel that has receptive fields with different sizes to further capture the contextual information and these features are fused to learn the primary rain streaks structures.Moreover,we illustrate that convolution layers at different scales have similar structure of rain streaks by statistical pixel histogram and they can be processed in the same operation.So,we deal with the rain streaks information at different scales by using multi-scale kernels de-raining layers with shared parameters,where we call this operation as multi-scale feature maps de-raining layer.Finally,we employ dense connections to connect multi-scale feature maps de-raining layers to maximize the information flow along features from different levels.4)Previous methods neglect the correlation between different layers with different receptive fields that loss a lot of important information.To better solve the problem,we develop a multilevel guided residual block that is the basis unit of our network.In this block,we utilize multilevel dilation convolutions to obtain different receptive fields and the layer with smaller receptive fields to guide the learning of larger receptive fields.Moreover,in order to reduce the model sizes,the parameters are shared among all multi-level guided residual blocks.Experiments illustrate that the guided learning improves the deraining performance and the shared parameters strategy is also feasible that provides basic for future small deraining models.5)Multi-scale has been applied in to many computer vision problems and achieves better performance.However,the correlation between different scales has not been explored in most methods.From this drawback,we present two types of scale-guide blocks and develop two combinations between the blocks to connect the correlation between different scales.One type of scale-guide is that small scale guides the large and anther is large scale guides the small.Moreover,we develop single-stage deraining method to recurrent networks,using LSTM to link every stage.