Generative Adversarial Network based on Resnet for Conditional Image Restoration
This work addresses image restoration for applications like photo enhancement, but it appears incremental as it builds on existing GAN and ResNet methods.
The paper tackles the problem of conditional image restoration from extremely degenerated images by proposing ResGAN, which embeds coarse images aligned to attributes as inputs and labels, and uses a ResNet-like straight path for direct transfer. Experimental results on MNIST, CIFAR10/100, and CELEBA datasets show higher accuracy compared to state-of-the-art GANs.
The GANs promote an adversarive game to approximate complex and jointed example probability. The networks driven by noise generate fake examples to approximate realistic data distributions. Later the conditional GAN merges prior-conditions as input in order to transfer attribute vectors to the corresponding data. However, the CGAN is not designed to deal with the high dimension conditions since indirect guide of the learning is inefficiency. In this paper, we proposed a network ResGAN to generate fine images in terms of extremely degenerated images. The coarse images aligned to attributes are embedded as the generator inputs and classifier labels. In generative network, a straight path similar to the Resnet is cohered to directly transfer the coarse images to the higher layers. And adversarial training is circularly implemented to prevent degeneration of the generated images. Experimental results of applying the ResGAN to datasets MNIST, CIFAR10/100 and CELEBA show its higher accuracy to the state-of-art GANs.