Super Resolution Image Reconstruction Based On The MRF-MAP Framework

Super-resolution reconstructiontechnology was developed by reconstructing ahigh resolution image by a series of low resolution under-sampled images, whichneeds to contain different information, and during the reconstruction process it ispossible to eliminate blur and noise of the image acquisition device, this technologycan be applied in many areas, and many people do research on this topic in recentyears.There are mainly three categories methods to obtain high resolution image,which is the interpolation algorithm, the reconstruction algorithm and thesuper-resolution reconstruction algorithm based on learning. The interpolationalgorithm is the earlier reconstruction algorithm,itrequiresthe input images havesub-pixel offset each other; the super-resolution reconstruction algorithm is a hotspotof current research, there areIBP (iterative back-projection)and POCS algorithm(POCS, projection onto convex sets) and so on. Learning methods has strictconstraints and limited range of applications. So these algorithms have someshortcomings, people are increasingly unable to meet the high demand.The pixel value of image area without edge change slowly, so one pixel has onlyrelationship with its neighborhood. The value of a point in an image can beconsidered implementation of MRF，it explores the relationship of pixel gray valuewith the gray value of its neighborhood. This article take the application of MRF inimage super-resolution reconstruction as prior acknowledge,based on based on theequivalence of MRF and GMR, use the energy function to represent the conditionalprobability of the MRF, which is make sure the MRF can be mapped to the globalsituation. To make a connection between one pixel and its neighbors through theMRF model, then obtain a global statistical result through Gibbs distribution. Theenergy function can be getting through Bayes framework, and MRF will be the priormodel. Analyzing the characteristics of Gauss MRF prior model and Huber MRF prior model, summarized some basic characteristics of the potential function and edgepreservation characteristics.According to the characteristics of potential function. Iuse an MRF model based on new potential function which can inhibit image noiseand keep the image edge detail, and so on.In this paper, simulated annealing algorithm can be used to solve thecombinatorial optimization problems, it can also be ascribed a stochastic optimizationalgorithm. There are some adjusted in this paper according the shortcomings oforiginal simulated annealing algorithm to make sure to get an global optimalresolution. It avoids a lot of unnecessary optimization process.