Research On Learning-Based Nonlinear Algorithm Of Face Image Super-Resolution

Image super resolution technology is a method that obtains high resolution image based on the signal-processing technology. By using a number of noisy, blurred low resolution images, It can produce a high resolution image through the signal-processing technology. So it can be applied in many important areas, such as video, imaging, remote sensing, medical, surveillance and military etc. In recent years, learning-based image super resolution technology has become more and more popular.By studying the characteristic of learning-based super resolution and the inherent property of face image, this paper proposes a non-linear face image super resolution algorithm based on semantic constraint patch-based Markov Network. The specific content of the thesis as follows:Firstly, this paper comprehensively reviews the basic concept and methods of image super resolution technology, and describes learning-based general image super resolution technology in detail. Then, this paper studies learning-based algorithm of single face image super resolution. Using Markov Network Model to express the reconstruction mechanism, by studying face image this special image, combining with the semantic characteristics of face images, and considering the differences between face images, and based on a aligned patch position constraint operation for searching the matched patch, this paper proposes a non-linear searching algorithm which decreases the complexity of the searching operation and increases the effect of matching and speeds up the convergence of the Markov network. After the matched high-resolution patches are collected, the algorithm uses these patches to dispose in vision compatibility, and directly output relative patches as restituted face image. The experimental result shows that the proposed algorithm has a good output quality, high efficiency and some practical value.