Block Markov Network Model-based Face Image Super-resolution Reconstruction Algorithm

The research of face image is one of hot questions in the field of the pattern recognition and computer graphics. Nowadays, the human face image resolution through the monitoring device is not effective and it's not so helpful to the human face recognition and the track and so on. The Supper Resolution (SR) technology is one kind of method which obtains the High Resolution (HR) image based on the signal processing technology. The basic theory of the SR is to put in the fusion of a sequence of low-resolution noisy blurred images and then to produce a higher-resolution image based on the signal processing technology. The SR has a very good application in human face research, long-distance image remote sensing, video frequency monitoring, and medicine domains and so on.The thesis focuses on the single human face image super resolution techniques with the goal aiming at an algorithm which is more simple, practical and suitable for real-time applications. This thesis comprehensively reviewes and narrates the SR technology concept, essential method and the SR algorithm. The key research lies on Learning-based SR algorithm. This thesis adopts Markov network (MN) model to propose a new frame description restructuring mechanism. We propose a novel algorithm that uses the location-restraint operation and the most compatible neighboring patches along horizontal dimension of the face to directly mosaic the high-resolution patches into the outcome. This method can reduce the complexity of search order, enhance the match relevance, speed up Markov network restraining, and simplify the implicit strata node computation. The experiment has been programmed by the VC++, and the human face image training sets in the experiment adopts 24 gradation images. The experimental result indicates the algorithm proposed in the thesis has some advantages such as high-quality output, high efficiency, which has some practical values.This thesis tries to develop a better and more intelligent learning-based algorithm, which can be helpful to improve the development of SR technology.