2018/07/30 by Adrian Bulat, Jing Yang, Bulat, Adrian +3 · 3 citations
Computer Science · Engineering · #Advanced Image Processing Techniques #Image Processing Techniques and Applications #Image and Signal Denoising Methods
paper · pdf · doi:10.48550/arxiv.1807.11458
This paper is on image and face super-resolution. The vast majority of prior\nwork for this problem focus on how to increase the resolution of low-resolution\nimages which are artificially generated by simple bilinear down-sampling (or in\na few cases by blurring followed by down-sampling).We show that such methods\nfail to produce good results when applied to real-world low-resolution, low\nquality images. To circumvent this problem, we propose a two-stage process\nwhich firstly trains a High-to-Low Generative Adversarial Network (GAN) to\nlearn how to degrade and downsample high-resolution images requiring, during\ntraining, only unpaired high and low-resolution images. Once this is achieved,\nthe output of this network is used to train a Low-to-High GAN for image\nsuper-resolution using this time paired low- and high-resolution images. Our\nmain result is that this network can be now used to efectively increase the\nquality of real-world low-resolution images. We have applied the proposed\npipeline for the problem of face super-resolution where we report large\nimprovement over baselines and prior work although the proposed method is\npotentially applicable to other object categories.\n