Fusion of noisy infrared and visible images based on anisotropic bivariate shrinkage

Abstract

The performance of wavelet-based image fusion can be improved by taking into account the statistical dependencies between wavelet coefficients. This paper focuses on the fusion of noisy infrared and visible images and presents a new image fusion approach based on the modeling of wavelet coefficients with an anisotropic bivariate Laplacian distribution function. The use of the anisotropic bivariate Laplacian model not only captures the dependencies between wavelet coefficients and their parents, but also fits the fact that the variances of wavelet coefficients of natural images are quite different from scale to scale. Based on this statistical model together with the assumption of additive white Gaussian noise, a closed-form anisotropic bivariate shrinkage function is derived within the framework of Bayesian denoising. Then an improved image fusion algorithm is proposed by incorporating the newly derived shrinkage approach into the fusion scheme. The performance of the proposed method was compared with three other methods and was found to perform the best in terms of both visual quality and objective metric Q AB / F n [17] when tested on four representative sets of infrared and visible images corrupted by additive white Gaussian noise.