Anisotropic total variation method for text image super-resolution

Abstract

This thesis paper addresses the problem of reconstructing a high resolution image from multiple low resolution text images. In general purpose image reconstruction area, Total Variation (TV) based regularization model is novel edge preserving and noise removing method for reconstructing high resolution image. Unfortunately, the total variation function was not efficient when it is applied for text image because of problems caused by specific properties of text images. Exposing those problems we proposed to solve them by using region complexity and stroke direction reflecting penalty function. We proposed an anisotropic total variation (ATV) function which reflects region complexity and stroke direction by involving position and direction dependent parameters to the total variation. And we presented a method to calculate those parameters directly from structure tensor field of initial image. Also we developed value reusing technique for faster implementation, which approximates calculation in homogeneous region to save computational cost during optimization. Experiments were performed with grayscale and color low resolution text images. Our method performed better with low computational cost than general purpose total variation in sense of minimizing Root Mean Square (RMS) error.