2021/01/14 by Valentin L. F. Wolf, Wolf, Valentin, Andreas Lugmayr +7 · 2 citations
Computer Science · #Advanced Image Processing Techniques #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #FOS: Electrical engineering #Image Enhancement Techniques #Image and Signal Denoising Methods #Image and Video Processing (eess.IV) #Machine Learning (cs.LG) #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.2101.05796
openalex publication_date 2021/01/14 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
The difficulty of obtaining paired data remains a major bottleneck for\nlearning image restoration and enhancement models for real-world applications.\nCurrent strategies aim to synthesize realistic training data by modeling noise\nand degradations that appear in real-world settings. We propose DeFlow, a\nmethod for learning stochastic image degradations from unpaired data. Our\napproach is based on a novel unpaired learning formulation for conditional\nnormalizing flows. We model the degradation process in the latent space of a\nshared flow encoder-decoder network. This allows us to learn the conditional\ndistribution of a noisy image given the clean input by solely minimizing the\nnegative log-likelihood of the marginal distributions. We validate our DeFlow\nformulation on the task of joint image restoration and super-resolution. The\nmodels trained with the synthetic data generated by DeFlow outperform previous\nlearnable approaches on three recent datasets. Code and trained models are\navailable at: https://github.com/volflow/DeFlow\n