Color face super-resolution reconstruction with higher-order singular value decomposition

Abstract

This dissertation proposes two novel frameworks for color face super-resolution reconstruction with higher-order singular value decomposition in four basic color systems such as RGB , YCbCr , HSV and CIELAB color system. The first framework is based on the linear regression model with MPCA since a color face image can be naturally described as tensors or multi-linear arrays. We find that the traditional method does not consider the correlation of data in each color channel. Therefore, there is an error in the face reconstruction process. In this dissertation, we investigate the performance of our proposed method in sense of effect of number of eigenvalue, effect of noise and complexity respectively and we can reconstruct the reasonable color face images which are compared with the ground truth color face images. In the second framework, we decompose each pair of low and high resolution training face images into a small patches and apply higher-order singular value decomposition in a tensor space. In color face reconstruction process, there are two steps : the first step tends to reconstruct a global face. Next step, the local detail is hallucinated from small overlapped patches. The experimental results from standard color facial database show that our second proposed framework can effectively reconstruct the color face images than the previous method. However, decomposing small patches in the training process will result in a more complicated process than that of the first framework.