Super-Resolution Algorithm Of Face Images Based On Markov Network

The term super-resolution (SR) is used to describe the process of obtaining a high-resolution (HR) image or a sequence of HR images from a set of low-resolution (LR) observations. Low-resolution is equivalent to low-frequency and high-resolution consists of high, middle and low frequency bands. There are in general two classes of super-resolution techniques: reconstruction-based (from input images alone) and learning-based (from other images). Of particular interest is face hallucination (which implies the high-frequency part of face image, must be purely fabricated), or learning high-resolution face images from low-resolution ones. Hallucinating faces is particularly challenging because people are so familiar with faces.Many multimedia applications rely on high-resolution images. Conventional methods employed several input low-resolution images to reconstruct the output high-resolution image. Since it is difficult to obtain sufficient numbers of input images, the effect of such methods is limited. Learning-based algorithms are popular SR techniques that use application dependent prior to infer the missing details in low resolution images. Learning-based algorithm under the framework of Markov Random Fields is emerging recently and attracts many researchers. However, most of such algorithms learn low-level knowledge about images. Recent researches indicate that learning high-level knowledge is also important and should be investigated further more. We follow this line and study the face image super resolution techniques based on image examples and Markov Network (MN).In this thesis, we investigate the face image super resolution technique or hallucination based on image examples and Markov Network (MN). We build a statistical model of high-resolution images and show how this model can be used to estimate a high-resolution image from a lower-resolution input image. Our goal is to make the algorithm of SR face images more simple, flexible and suitable for real-time applications. Under the framework of standard MN, we propose a novel algorithm that uses the location-restraint operation to increase the probability value of observation function in MN. For hidden nodes of MN, we use the most compatible neighbouring patches among k closest ones that are selected by combining the pixel colour feature matching and location restraint operation. This can increase the transition function of MN and works very well even under the condition where the alignment of the images in the training dataset is not accurate. By these measures, we generate an adaptive MN. Achieving accurate, fast and a simpler, flexible and suitable algorithm for realtime application of the input images is a critical step in super-resolution processing. Motivated by this basic requirement, the techniques presented in this thesis are tested in practical experiments. We carry out related experiments with the image database and demonstrate the effectiveness and performance of the proposed algorithms. Due to the simplicity of computing MN, our algorithm is practical and suitable for real-time applications. We find that location restraint mechanism can potentially increase the probability value of observation function in MN. Especially, with a small training dataset, it extremely improves the performance of searching the closest patches.