Swarm selection method for multilevel thresholding image segmentation

Multilevel thresholding is one of the most popular approaches used for image segmentation. Several methods have been used to find the threshold values; however, metaheuristic (MH) methods achieve better results than the other classical methods. Nevertheless, each metaheuristic method has its limitation such as high computation; moreover, it can suffer from premature convergence when the level of the threshold increases. To solve this problem, there is a trend to hybridize the MH algorithms together; however, there are several MH algorithms that can be combined. Consequently, we need to seriously determine which combination and number of such algorithms must be combined. Therefore, this paper proposes an intelligent framework that provides experts with a tool for solving the problem through selecting a suitable number of swarm algorithms from eleven such algorithms. The proposed method, which is called swarm selection (SS), distributes the eleven algorithms into two groups, the first of which is called the control group and only contains the differential evolution (DE) algorithm. The aim of the DE algorithm in the control group is to determine the best combination of the other ten algorithms in the second group. Meanwhile, the selected algorithms in the second group work together to find the optimal threshold values that maximize the Otsu function. A series of experiments have been performed on six test images using nine threshold levels. The performance of the proposed SS method is compared with the other three hybrid algorithms and two non-hybrid algorithms. The experimental results demonstrate that the proposed SS method outperforms the other image segmentation methods in terms of the performance measures, peak signal-to-noise ratio (PSNR), structure similarity index (SSIM), fitness function value and computational time. In addition, by using this intelligent framework, experts performing applications that depend on image segmentation will save computational time in identifying suitable methods. (C) 2019 Elsevier Ltd. All rights reserved.