Siamese Network for Single Image Rain Removal

Abstract

This work focuses on removing rain streaks from the single image using the Siamese network. The Siamese network with an unique structure, where the inputs are paired images, are utilized to obtain the discriminative features of the input. We explore a method using the Siamese neural network to learn the difference between rainy images and clean images. The convolutional architecture is used for feature learning. Once the network has been tuned, the key feature of the rain is learned, then the rain streaks can be removed. By training the network on synthetic data, our framework can be generalized in the real rainy image tests. Numerical experiments show that the proposed method performs well in removing rain for the synthetic rainy images in terms of visual and quantitative measures.