Matrix Based Regression with Local Position-Patch and Nonlocal Similarity for Face Hallucination

摘要

Learning based face hallucination methods have received much attention and progress in past decades. As opposed to the existing methods, where the input image patch matrix is first stacked into vectors before combination coefficients calculation, this paper directly uses the matrix based regression model for combination coefficients computation to preserve the essential structural information of the input matrix. For each low-resolution local patch matrix, its combination coefficients over the same position image patch matrices in training images can be computed. Then the desired high-resolution patch matrix can be obtained by replacing the low-resolution training samples with corresponding high-resolution counterparts. The nonlocal self-similarities are finally utilized to further improve the hallucination performance. Experiments conducted on the FERET face dataset indicate that our method could outperform other state-of-the-art algorithms in terms of both vision and quantity.