Hybrid reconstruction in Bayesian domain

Abstract

Image reconstruction in Bayesian framework is far more advantageous over other reconstruction methods like convolution back projection, weighted least square method and maximum likelihood estimation. The power of Bayesian estimation ties in its ability to incorporate the prior distribution knowledge, enabling better reconstruction. Proper specification of clique potentials in Bayesian estimation plays a crucial role in the reconstruction process by favors the presence of desired characteristics in the image lattice like nearest neighbor interactions and homogeneity. Homogenous Markov random fields have been successfully used for modeling such interactions. Though reconstructions produced by such models are far more efficient, they often require large iterations for producing an approximate reconstruction. To deal with this problem, we have extended the Bayesian estimation in order to support sharp reconstruction. We propose to use sharp potential in Bayesian estimation once an approximate reconstruction is available using homogenous potentials in Bayesian domain The advantage of the proposed potential is its ability to recognize correlated nearest neighbors. The proposed reconstruction is a hybrid of both smooth and sharp potential in Bayesian framework and hence it is termed as hybrid reconstruction. Simulated experiments have shown that the proposed hybrid estimation method produces superior and sharp reconstruction as compared to the reconstruction produced by other Bayesian estimation methods.