Multi-detail convolutional neural networks for single image rain removal

This paper, proposes a single image rain removal method based on multi-detail convolutional neural network. Considering that the raindrop information is mostly presented in the high frequency part of the rainy-image, the proposed method exploits the guided filter to decompose the rainy-image into a smooth image and multi-detail images with different frequency distribution, and then a multi-detail convolutional neural network is developed to learn the mapping relationship between the rainy-image and the rainless image, so as to obtain a rainless image. Considering that it is difficult to collect rainy-images and rainless images in the same scene in reality, we use rainless images and artificially synthesized rainy-images as training data, while utilize the synthesized rainy-images and real rainy-images as the testing data. The experimental results have shown that the proposed method can effectively remove the raindrop information from the rainy-image.