Residual Adaptive Mask Generative Adversarial Network for Image Raindrop Removal

Abstract

Single image raindrop removal is an extremely challenging task since the raindrop regions of various shapes and sizes are not given and the background information of the occluded regions is completely lost for most part. In this paper, a novel two-stage residual adaptive mask generative adversarial network (RAM-GAN) is developed for single image raindrop removal, in which the raindrop regions can be automatically detected and a restored image without raindrops is generated. Moreover, the residual adaptive mask block (RAMB) structures and residual dense adaptive mask modules (RDAMM) are proposed to be the main components constructing the network. The proposed RAMB structure can serve as a feature selector which adaptively enhances the effective information and suppress the invalid information. Each block is processed into two branches: soft mask branch and trunk branch. A mask is generated by the soft mask branch to softly weigh the features processed by the trunk branch. In addition, RDAMM, the residual densely connected module based on RAMB structure, is proposed to maximize the information flow among different blocks and guarantee better convergence. Our experimental results have demonstrated that our method can effectively remove raindrops while well preserving the image details, which outperforms the state-of-the-art methods quantitatively and qualitatively.