Bayesian Estimation of Global and Local Variances for Improved Perceptual Quality of Fused Medical Images

This paper presents a novel image fusion method particularly effective in increasing perceived image quality. Such capability is achieved by reducing the natural distribution of the input image pixels in order to enable the detection of finer details, a goal largely missed by existing image fusion techniques. The procedure entails minimizing the local variance of input image pixels by combining Bayesian estimation with a greedy-iterative algorithm to preserve variations in local pixel neighborhood. The global area mean of the input image pixels is also minimized by using Bayesian estimation to control pixel gradient and mean values in order to compensate for illumination inhomogeneities. Bayesian estimation of pixel mean and variance, both locally and globally, helps increase correlation between pixels, thereby preserving the finer image details during fusion. The final images are fused and objectively evaluated based on entropy, standard deviation and root-mean-squared error, showing results that outperform existing image fusion methods.