Image magnification based on adaptive MRF model parameter estimation

The Markov random field (MRF) model, whose model parameters specify the amount of smoothness in an image, is a popular approach to image magnification. The model parameters must be estimated accurately in order to obtain an elegant solution. The conventional parameter estimation methods consider an image to be homogeneous and have a high computational complexity. However, images are usually not homogenous; using only one set of parameters cannot describe a whole image effectively. We therefore devise an adaptive parameter estimation method for the MRF model to reduce the blocky artifact while preserving the edges in the (high-resolution) HR image. In our method, an initial estimated HR image is divided into small blocks, and the respective parameters are then estimated. Their values are defined as inversely proportional to their energy in the corresponding direction. Then, the gradient descent algorithm is employed iteratively to obtain an improved HR image in a Bayesian MAP framework. Experimental results show that, when compared to the MRF model with a fixed set of parameters, using the MRF model with our adaptive parameter estimation method can produce a magnified image with the edges and texture well preserved. Both the PSNR and visual quality of our proposed method are much better than the fixed-parameter method.