DD-GAN: pedestrian image inpainting with simultaneous tone correction

Yuelong Li, Tongshun Zhang, Junyu Bi, Jianming Wang

Abstract: Accompanied by the rapid popularization of camera surveillance devices, tremendous pedestrian images can be acquired. Since a huge part of surveillance cameras are deployed out-of-door, many obtained images may contain some kinds of flaws, such as content defection and tone disorder. In this paper, a newly designed adversarial framework, namely DD-GAN, is proposed which could simultaneously recover both mentioned image damages. DD-GAN consists of three parts: generator, tone discriminator and inpainting discriminator. DD-GAN emphasizes the fusion of image inpainting and tone correction through the GAN network. In this way, we construct two generative adversarial losses, for two completely different functions, achieving the efficient combination of tone correction and image inpainting. Extensive experiments are conducted to verified the effectiveness of the proposed approach.