A Novel Supervised Bi-Level Thresholding Technique Based on Particle Swarm Optimization

Abstract

Thresholding is an important step in many computer vision tasks such as image segmentation. However, selection of the optimal threshold value often appears as a challenging problem to researchers. In this paper, a novel approach for determining of the optimal threshold value is proposed which works based on supervised version of particle swarm optimization (PSO) method. Our proposed technique employs the ability of Otsu method for minimizing within-class variance as well as a preprocessing step aimed at transferring maximum region uniformity to output image. In preprocessing stage, we attempt to maintain as much information as possible by obtaining a Canonical image which includes the most visually important features of the original image. Afterwards, the PSO algorithm is applied to find the optimal threshold value. The PSO method chooses the initial values by exploiting Otsu approach along with an objective function which maximizes the similarity to the Canonical image. The objective function utilizes the sequential connected component labelling algorithm. The presented approach has been tested against standard datasets for comparison with well-known thresholding methods in the literature. Experimental results demonstrate the superiority of our algorithm to other methods in both subjective and objective evaluations.