Text Image Super-resolution Recovery

The task of Super-Resolution(SR) image restoration is to get higher resolution image from several low-resolution image sequences of the same sence. This technique takes advantage of not only inner image pixel correlations but also available imformations between frames in the sequences. Compared with traditional image restoration techniques, it can reconstruct more detail infromations of the image.As a special kind of images, text iamge has its own characters . First, the typical distribution of a text image histogarm contains two peaks ,that is the histogram of a text image is bimodal. Second, in structure, the frontground of text image is constructed by characters with limited structures. Take advantage of it we can get better algorithms special for text image super-resolution restoration.This paper firstly presents some of the background theory of SR, and then an extensive survey of the literature on multiframe super-resolution restoration is presented. The main task of the paper is some improve in the algorithm special for text image super-resolution restoration in both stochastic reconstruction and total variation reconstruction framework.In stochastic reconstruction framework, maximum likelihood (ML) estimater and maximum a posterior (MAP) estimater was introduced. For maximum a posterior estimater we used Huber penalize function and Bimodal restriction as the a-priori regulizer. In this part, the main contribution is the analysis of the Bimodal restriction, we find essentially it's a special auto-adaptive histogram modification.In total variation estimater ,we introduced the bilateral filter, and showed that a single iteration of the weighted robust restoration yields the bilateral filter, and explained the bilateral filter has the edge-preserving character. We used the biliter filter theory to total variation estimate, consider both the spatial and pixel value's contribution to weights, we realized the biliter- total variation reconstruction algorithm. Further more, improvements consider the construction of Chinese character was done to the regulizer. And experiment results proves it's better in Chinese character text image super-resolution resconstruction.