Blur, Noise, and Compression Robust Generative Adversarial Networks

2 RIKEN Real (a) Training images Generated (FID: 34.9) (b) GAN (baseline) (c) BNCR-GAN (proposed) Generated (FID: 24.2) Blur + Noise + Compression Figure 1. Examples of blur, noise, and compression robust image generation. Although recent GANs have shown remarkable results in image reproduction, they can recreate training images faithfully (b), despite degradation by blur, noise, and compression (a). To address this limitation, we propose blur, noise, and compression robust GAN (BNCR-GAN), which can learn to generate clean images (c) even when trained with degraded images (a) and without knowledge of degradation parameters (e.g., blur kernel types, noise amounts, or quality factor values). The project page is available at https://takuhirok.github.io/BNCR-GAN/ .