Gradient based evolution strategy for parametric illumination correction

Abstract

This chapter deals with the issue of illumination inhomogeneity correction in images. The approach followed is that of estimating the illumination bias as a parametric model. The model is a linear combination of Legendre polynomials in the 2D or 3D space. The estimated bias is, therefore, a smooth function characterized by a small set of parameters that define a search space of lower dimension than the images. Our work is an enhancement of the PABIC algorithm, using gradient information in the mutation operator hence we name it GradPABIC. We apply our algorithm, the PABIC, and a conventional Evolution Strategy (ES) over a set of synthetic images to evaluate them, through the comparison of the correlation between the recovered images and the original one. The PABIC and the EE are allowed the same number of fitness computations, while the Grad PABIC number of fitness evaluations is two orders of magnitude lower, because of the gradient computation added complexity. Finally, we present some results on slices of a synthetic MRI volume.