Abstract
Recently, generative adversarial networks (GANs), which learn datadistributions through adversarial training, have gained special attention owingto their high image reproduction ability. However, one limitation of standardGANs is that they recreate training images faithfully despite image degradationcharacteristics such as blur, noise, and compression. To remedy this, weaddress the problem of blur, noise, and compression robust image generation.Our objective is to learn a non-degraded image generator directly from degradedimages without prior knowledge of image degradation. The recently proposednoise robust GAN (NR-GAN) already provides a solution to the problem of noisedegradation. Therefore, we first focus on blur and compression degradations. Wepropose blur robust GAN (BR-GAN) and compression robust GAN (CR-GAN), whichlearn a kernel generator and quality factor generator, respectively, withnon-degraded image generators. Owing to the irreversible blur and compressioncharacteristics, adjusting their strengths is non-trivial. Therefore, weincorporate switching architectures that can adapt the strengths in adata-driven manner. Based on BR-GAN, NR-GAN, and CR-GAN, we further proposeblur, noise, and compression robust GAN (BNCR-GAN), which unifies these threemodels into a single model with additionally introduced adaptive consistencylosses that suppress the uncertainty caused by the combination. We providebenchmark scores through large-scale comparative studies on CIFAR-10 and agenerality analysis on FFHQ dataset.