Research On Character Oriented Image Super-Resolution Method

Improving the image resolution to obtain more available information is a research hotspot in image processing. This dissertation follows the application demand of the intelligence analysis, deals with single-frame image to raise the resolution, and focuses on an identify-oriented super-resolution approach basing on specific object (character) in the image. In this paper, algorithm contains four steps. First, the target region is located in order to reduce the calculation of the volume. The second, edge and ridge is detected. Third, the edge and ridge is interpolated using adaptive interpolation method. Fourth, interpolating results is optimized by using of image quality evaluation function.The main contents are summarized as below:1. Image super-resolution methods are reviewed and the representative, stimulating super-resolution algorithms are analysised deeply.2. A pre-locating method based on two-class dissimilarity measure and character relevancy is proposed. The region of interest is confirmed by preliminary filtering which contains two steps: first the plain region is removed; and then the character region is found in the remaining region by adaptive gray-scale segmentation and characteristics-based character positioning.3. A ridge detection method basing on symmetric gradient direction is proposed aiming at the disadvantage of common methods in existence. The point is dectected whether on the ridge line according to the symmetric gradient direction between its neighborhood pixels, and then the direction of the ridge line is determined.4. A adaptive interpolation method for character feature basing on sliding-window is proposed. This method integrates the advantage of the direction-based interpolation algorithm and the correlation-coefficient-based interpolation algorithm which has a better performance.5. An optimization processing to interpolation results is implemented by two-class measure and line-based measure as evaluating standards. The experiment results validated that the methods advanced in this dissertation is better than the common methods.