A novel strategy of differential evolution algorithm's crossover operator based on graylevel clusters similarity for automatic multilevel image thresholding

Automatic multilevel image thresholding approach has wide optimal solution search space due to its ability to determine thresholds number and positions, simultaneously. Searching its optimal solution using standard Differential Evolution (DE) algorithm can decrease its efficiency due to slow convergence. Therefore, a strategy that can restrict the search space is needed in order for optimizations being efficient. In this paper we propose a novel strategy of DE's crossover operator based on graylevel clusters similarity for automatic multilevel image thresholding. We restrict the search space by only recombining graylevel clusters which have small similarity. Graylevel clusters similarity is performed by computing the inter-class and intra-class variance of adjacent graylevel clusters. Experiments on grayscale image of Berkeley Segmentation Dataset (BSDS500) show that the proposed crossover strategy can generate segmented images with misclassification error of 7.96% better than those of existing crossover strategies which not compute the graylevel clusters similarity. It only requires the average of 636 generation to find the optimal solution less than compared crossover strategies.