Super-resolution texture synthesis using stochastic PAR/NL model

Abstract

Super-resolution texture synthesis using a locally-adaptive stochastic signal model is investigated in this work. The 2D random texture is modeled by a piecewise auto-regressive (PAR) process whose parameters are determined by a non-local (NL) training procedure and, consequently, it is called the PAR/NL model. Unlike previous work that applies the NL scheme to image pixels directly, the proposed PAR/NL scheme applies the NL scheme to PAR model parameters by assuming that these parameters are self-similar. Furthermore, we describe a probabilistic method for PAR/NL model computation using the maximum a posteriori (MAP) and the expectation–maximization (EM) principles. Experimental results are given to demonstrate the synthesis performance of the proposed PAR/NL technique, which can boost texture detail and eliminate the blurring artifact perceptually.