Pixel-wise dictionary learning based locality-constrained representation for noise robust face hallucination

Abstract

Recently, the sparsity and locality constrained linear coding (SLLC) attracted much attention in image super-resolution (SR) applications. However, the current SLLC based SR methods are too weak to handle the impulse noise problem. Therefore, this work presents a dictionary learning-based locality-constrained representation (DLcR) for robust face hallucination. It simultaneously hallucinates the face images and suppresses the noise (outliers) using the noisy pixel information in the observed faces. It first identifies the correct position of the noisy pixels in the input face image and then learns the position-wise low-resolution (LR) dictionary images with the detected noise. This learning makes the similar noisy structure of LR dictionary images as the input LR face, which minimizes the error in optimal weight reconstruction. In addition, the performance of DLcR is improved further by imposing the appropriate thresholds on the LR dictionary, named T-DLcR. The thresholding of the LR dictionary in T-DLcR leads to represent the input LR patch through its nearest LR dictionary patches with the precise reconstruction weights. The comparison results of T-DLcR with several position-patch based face SR methods and recent deep learning-based SR method show its superiority on standard as well as real-world face datasets.