Single Image Deraining Based On Adaptive Perceptual Network

There are some problems such as rain streaks occlusion and fuzzy background details in rainy scenes,which lead to the degradation of collected image and seriously affect the performance of outdoor computer vision system.In this paper,single-image deraining based on deep learning methods is deeply researched and two novel deraining methods are proposed for residual rain streaks,serious detail loss and color distortion in existing methods.The main research work is summarized as follows:To avoid residual rain streaks and serious details loss in restored image caused by incomplete extraction of image features,a single-image deraining method based on adaptive perceptual pyramid network is proposed.It builds a multi-scale feature pyramid based on wavelet transform as the reference network.Each scale uses an adaptive rain streak perceptual sub-network for feature extraction and detail restoration with progressive rain removal between adjacent scales.In each sub-network,through skip-connection to extract shallow layer features to the deeper layers for reusing and guiding the extraction of deep features.The adaptive rain streak perceptual block uses non-local operation and shared dilated convolution to expand receptive fields and rain removal adaptively through the attention mechanism.Extensive experimental results on synthetic and real datasets demonstrate that the proposed method effectively removes rain streaks with preserving vivid image details.Most of the existing methods take into no account the diversity of rain streaks in real world,which results in incomplete removal of rain streaks in different directions and color distortion.Thus,we propose a residual deraining network based on direction guidance which is used to make full use of spatial context information.The network is constructed based on encoder-decoder network with residual encoder branch.Meanwhile,a dynamic convolution kernel selection strategy is incorporated into the feature extraction sub-network,which captures and recursively removes the rain streaks in different receptive fields.To further improve visual quality of the restored image,a color refine module is designed to compensate for the color information of the restored image.Experimental results show that the proposed method recovers high quality rainfree images on synthetic and real datasets.In addition,the restored images contain rich colors and details with ideal visual effect.