S. M. Mostafavi - Sabzevar Tarbiat Moallem University
J Haddadnia - Sabzevar Tarbiat Moallem University
P Moallem - University of Isfahan

This paper is concerned with a practical face super resolving (hallucinating) approach. The image reconstruction is based on high and low resolution image pairs. A patch is removed from the low resolution input image and is compared with the low resolution training images patches. By this comparison a weight for each low resolution images patch is gained. After estimating the corresponding weights, the learning databases high resolution images patch which is a pair for the low resolution images patch is used to hallucinate the input low resolution image. The high resolution output image is constructed by integrating the high resolution hallucinated patches. Unlike neither statistical nor manifold approaches this method uses the exact available data in its learning database to hallucinate the output image. Experimental results show that the proposed method generates higher quality hallucinated images demonstrating performance superiority and novelty in terms of single modal face hallucination

Face hallucination, Face Super-resolution, Image Patches