Parameter optimization for non-local de-noising using Elite GA

Non-local means de-noising is a simple but effective image restoration method. It exploits usual redundancy found in real-life images. It computes similarity between patches of pixels, in a non-local window, instead of pixels themselves. This similarity measure defines participation/weight of each pixel in the de-noising process. In this research study, non-local means de-noising has been applied to noisy synthetic and brain MR images by optimizing its parameters through Genetic Algorithm. Elite Genetic Algorithm, a novel idea, has also been proposed to optimize several parameters of the non-local framework. It works in a hierarchical structure i.e. K Primary GAs and one Secondary GA. Each Primary GA evolves with independent population and gives rise to n_k elite chromosomes after t generations, which collectively serve as population of Secondary GA. Evolution with Elite GA results in improved speed of convergence as Secondary GA starts its evolution with more fit chromosomes instead of randomly generated population. These elite chromosomes are expected to be better solutions, thus have higher probability to approach global minima/maxima in no time. Algorithm has been tested on the said images and improved convergence rate has been observed for Elite GA. Moreover, the individuals selected by Elite GA are as fit as traditional GA as verified by PSNR and RMSE results.