Face Super Resolution by Patch-Based Sparse Coding

This paper proposes a new method for improving face image resolution (i.e. face super-resolution method). Our method utilizes a single high-resolution sample face image to infer the lost information in low-resolution test face image while all the test face image and sample face image are divided into image patches as atoms by sliding windows sampling. The entire high-resolution sample face image patches are formed into a raw dictionary and then compacted and optimized by K-SVD algorithm to get a trained high-resolution dictionary. Correspondingly, we can obtain a low-resolution dictionary by down sampling each atom and any low-resolution patch of test image can be linearly represented by the low-resolution dictionary with sparse coding method. With the one-to-one correspondence property for high-resolution dictionary and low-resolution dictionary, we can reconstruct a high-resolution test face image by mapping the atoms and coefficients. We conduct a series of experiments to present our method which can achieve satisfactory results and be ready for practical use.