Learning-Based Super-Resolution System Using Single Facial Image and Multi-resolution Wavelet Synthesis

Abstract

A learning-based super-resolution system consisting of training and synthesis processes is presented. In the proposed system, a multi-resolution wavelet approach is applied to carry out the robust synthesis of both the global geometric structure and the local high-frequency detailed features of a facial image. In the training process, the input image is transformed into a series of images of increasingly lower resolution using the Haar discrete wavelet transform (DWT). The images at each resolution level are divided into patches, which are then projected onto an eigenspace to derive the corresponding projection weight vectors. In the synthesis process, a low-resolution input image is divided into patches, which are then projected onto the same eigenspace as that used in the training process. Modeling the resulting projection weight vectors as a Markov network, the maximum a posteriori (MAP) estimation approach is then applied to identity the best-matching patches with which to reconstruct the image at a higher level of resolution. The experimental results demonstrate that the proposed reconstruction system yields better results than the bi-cubic spline interpolation method.