In order to achieve better derainin

In order to achieve better deraining effects, inspired by the work done by He K et al in[33], this article uses a method called gudied filter to train the network. The model is expressed as follows, By leveraing prior image processing knowledge and experimental results, after applying loss-pass filter such as [34,35] to the image, a base image can be obtained, which is similar to the base image of original, IBase ≈ JBase. The image obtained by subtracting the original image from the base image is a detail view of original image. The detial image of both input image and target image is shown below Fig. Detial images can be utilized in training since it reserves the most of the information while dereases the computing complexity. Trained detail image can be added to the base image to get the de-rained images. Training on the detail images is potent due to several reasons. Firstly the detail image is sparser comparing to the original images from pixel wise, since most of the area is close to zero. Furthermore, by leveraging the sparse characteristic of detail images, the convolution neural networks could get convergence faster and easier to some extent.