Researches On Text Image Super-resolution Algorithm

Super-resolution aims to obtain high resolution images from the low resolutionversions. It is an economical way to improve the quality of image and has become oneof the hottest issues in the image processing field recently. This paper focuses onsuper-resolution for single text image. Different from natural image, the text image isusually piecewise smooth and contains lots of strong edges. So at first, conventionalsuper-resolution methods for natural images are introduced and their performances ontext images are tested by experiments. Then two different methods are proposedconsidering the characteristics of text images. An adaptive manifolds filter basedsuper-resolution method is developed for gray scale text images; and a real-timematting based method is designed for color text images. The outcomes of this workare as follows:1. An adaptive manifolds filter based super-resolution algorithm for gray scaletext image is proposed: Firstly the text image is divided into detail layer and baselayer by using the adaptive manifolds filter. Then the detail layer is improved by anadaptive enhancement algorithm. The improved detail layer is merged with base layer,then the combined image is interpolated by the bicubic method. At last joint filteringis performed to remove the background of the interpolated text image. The proposedsuper-resolution algorithm is computationally efficient. It can effectively enhance thetext edges even with the present of noise. The proposed method can improve visualeffect and accuracy of text recognition.2. A real-time matting based super-resolution algorithm for color text image isproposed: Firstly, a guided filter based real-time matting algorithm is designed todecompose input color text image into three components: text edge layer, foregroundcolor and background color layers. Considering the different characteristics of thethree layers, the less informative foreground and background layers are upsampled bythe bicubic interpolation while the text edge layer is upsampled by an edge-enhancedinterpolation. Finally, the three upsampled layers are composed to obtain the outputhigh resolution color text image.3. Software of the super resolution algorithm for text image is written: Theaforementioned super-resolution algorithms are implemented with C++. A graphicuser interface is developed to make the algorithms easier to use. With this software, users can load their own images as input, convert color image to gray scale one,modify the values of parameters, and save the super-resolved images.In the end, we summarize the work of this paper and the further research ispointed out.