Hallucinating Very Low-Resolution Unaligned and Noisy Face Images by Transformative Discriminative Autoencoders

摘要: Most of the conventional face hallucination methods assume input image is sufficiently large and aligned, all require to be noise-free. Their performance degrades drastically if tiny, unaligned, contaminated by noise. In this paper, we introduce a novel transformative discriminative autoencoder 8X super-resolve unaligned noisy tiny (16X16) low-resolution images. contrast encoder-decoder based autoencoders, our method uses decoder-encoder-decoder networks. We first employ decoder network upsample denoise simultaneously. Then use encoder project intermediate HR faces aligned noise-free LR faces. Finally, second generate hallucinated Our extensive evaluations on very dataset show that achieves superior results outperforms state-of-the-art margin 1.82dB PSNR.