Quality Aware Generative Adversarial Networks

Generative Adversarial Networks (GANs) have become a very popular tool forimplicitly learning high-dimensional probability distributions. Severalimprovements have been made to the original GAN formulation to address some ofits shortcomings like mode collapse, convergence issues, entanglement, poorvisual quality etc. While a significant effort has been directed towardsimproving the visual quality of images generated by GANs, it is rathersurprising that objective image quality metrics have neither been employed ascost functions nor as regularizers in GAN objective functions. In this work, weshow how a distance metric that is a variant of the Structural SIMilarity(SSIM) index (a popular full-reference image quality assessment algorithm), anda novel quality aware discriminator gradient penalty function that is inspiredby the Natural Image Quality Evaluator (NIQE, a popular no-reference imagequality assessment algorithm) can each be used as excellent regularizers forGAN objective functions. Specifically, we demonstrate state-of-the-artperformance using the Wasserstein GAN gradient penalty (WGAN-GP) framework overCIFAR-10, STL10 and CelebA datasets.