Example-Based Face Image Super-Resolution Taking Into Consideration Correspondence Of Facial Parts

Example-based methods are widely adopted in image super-resolution (SR) to generate clear, high-resolution (HR) images from low-resolution (LR) images. This paper reports our study of example-based algorithms on LR facial images that exploit the relationship between LR and HR image patch pairs in a database using a Markov random field (MRF) model. We aimed to restore each part in a face with patches related to the parts. We generated patches with information on their original positions from a set of normalized facial images where their facial feature points were approximately in the same position. The nearer candidate patch's position to the target position yielded higher compatibility of the facial parts. Our algorithm restored LR images by combining the proposed facial parts' compatibility function with the conventional function to find the best set of estimated HR patches. The final SR results were obtained by stitching the inferred HR patches from an iteration process. An experiment on a set of facial images demonstrates that the combined compatibility function achieves the best quality in the resulting image, i.e. 30.39 [dB], in terms of the peak signal-to-noise ratio (PSNR) compared to the previously achieved quality of 29.65 [dB]. (C) 2017 Institute of Electrical Engineers of Japan. Published by John Wiley & Sons, Inc.