Multi-level Guided Residual Network For Single Image Deraining

Under rainy conditions,affected by rain lines or rain bars,the captured images will be degraded,information will be destroyed and lost,the visual visibility of the images will decrease,and the image will be blurred or distorted.In real-world work tasks,high-quality images are often required,and degraded images limit the outdoor advanced visual processing algorithms,such as target detection,image segmentation and other applications.Therefore,it is an urgent problem to develop an effective image removal algorithm to recover rain-free images from rainy images.As the bottom processing task in the computer vision system,image rain removal has very important research value in the current stage.However,the traditional image rain removal method based on low-rank representation,sparse representation and Gaussian mixture model is to make some prior assumptions on the rain map based on the statistical characteristics of rain,but these artificial priors are often only applicable to specific Circumstances do not reflect the characteristics of rain in complex scenes,and have certain limitations,so the effect is not good in actual image rain removal tasks.In recent years,a large number of deep learning methods have been proposed to learn the depth characteristics of rain bar information from massive data,which can avoid formulating optimization models and manual design priors.However,most of the existing deep learning-based image rain removal methods are The removal of rain lines will also remove the texture details of areas without rain,resulting in an overly smooth background and color distortion.In addition,many current methods often ignore the correlation between features of different scales,which is very important for learning the information of rain bars in images.In order to better learn the characteristics of the rain bars in the image and help solve the problem of image removal,this paper proposes an effective single-frame image removal algorithm based on multi-scale guided residual network based on the physical model of image removal.On the one hand,the multi-scale guided residual network structure uses expansion rates with different exponential levels to obtain different spatial texture feature information.On the other hand,this paper uses dense connections to enhance the internal connections between features of different scales in the network.Through the use of guided learning,the feature layer of the smaller receptive field guides the larger receptive field of the adjacent layer,so that the entire network structure has a more powerful rain bar feature representation ability,and it is better to learn different types of rain at different locations in the image.This piece of information is conducive to recovering the rainless image from the rainy image.This paper has conducted a large number of quantitative and qualitative experiments on multiple public benchmark data sets.Compared with the latest methods,the experimental results of the method proposed in this paper have higher numerical indicators and better visual effects.At the same time,it has been tested.Compared with the latest method,the running time of a 512*512 image has higher computational efficiency.