Methods for gamma invariant colour image processing

摘要

This article presents methods for normalizing natural RGB images with respect to the group of gamma adjustments. Applications of the normalization include image enhancement and gamma invariant indexing. By utilizing the logarithmic domain it is possible to define both histogram-based and spatially based normalization methods involving operations that commute with gamma, which is an essential benefit in practical algorithms. The normalization can be refined using a neural network or other empirically optimized system in a way consistent with the normalization principle. It is also possible to perform normalization simultaneously with respect to gamma and linear scaling. Four algorithms were tested using a set of over 3600 images. The average ratio between the computed gamma values and subjective optimum gammas was less than 1.3 for the best algorithm, which utilized a neural network. The gamma invariance of the algorithms and their stability under perturbations were good except in the presence of zero values, at which the logarithm is singular.