Super-resolution aims to produce a high-resolution image from a sequence of low-resolution images. Bayesian super-resolution is a stochastic approach for super-resolution. The Bayesian super-resolution framework provides a chance to utilize the prior knowledge over the high-resolution images. There were many previous works for the image prior model. However, almost of them were considered for the general images.
In this thesis, we present a new image prior model appropriate to text images in Bayesian super-resolution framework based on training example images. Two basic ideas are considered in the proposed prior model. One is to obtain extra information over the high-resolution image from an underlying high-resolution image. The underlying high-resolution image comes from training examples. Training examples are composed of pairs of high-resolution training images and their blurred training images. And the other is to model the text image property that a text image is composed of two homogeneous regions of the text region and the background region. Selective smoothing using an edge map performs local smoothing in each region. These two ideas are modeled in Markov random field as a clique system and its energy function.
Experiments are performed with scanned images from journal and Korean book. Our prior model shows significantly improved results over other prior models for general images and a text-specific prior model in sense of better looking and minimizing RMS error. And binarized RMS error also shows that the proposed prior model is more appropriate to text images than other prior models.