| Summary: | Segmentation is a key process in image analysis as it involves dividing and separating a digital image into different regions of pixels for further detailed investigation. Among others, multilevel thresholding is a robust and most widely adopted image segmentation approach. To find the optimal multilevel threshold values, various heuristic and meta-heuristic algorithms have been applied to segment COVID-19 medical images. Although effective, these algorithms get stuck in local optima and need proper parameter tuning for solving optimisation problems. To address these issues, this work proposes a novel multilevel thresholding strategy using Exponential Monte Carlo with Counter (EMCQ) hyper-heuristic for the segmentation of Computed Tomography (CT) chest images. EMCQ uses four low-level heuristic sets adopted from the teaching learning-based optimisation (TLBO) algorithm, flower pollination algorithm (FPA), genetic algorithm (GA), and Jaya algorithm. One of these low-level heuristics is selected by EMCQ in each iteration based on the best performance. Apart from the best low-level heuristic, EMCQ also gives a chance to low-performing heuristics to search for the optimal threshold values using their probability density function. For performance evolution, common image processing performance metrics like the feature similarity index method (FSIM), mean square error (MSE), peak signal to noise ratio (PSNR), and structural similarity index measure (SSIM) are utilised. Experimental results of the proposed strategy are compared with other advanced meta-heuristic algorithms using the Otsu and Kapur fitness functions. For the CT images, the proposed strategy achieved mean best values of 83% for PSNR, 95% for MSE, and 66% for FSIM. On the classical image dataset, it outperformed an existing hyper-heuristic strategy by obtaining 70% best SSIM values compared to 50% from the existing approach. The validity of the proposed strategy is further confirmed through statistical analysis using the Wilcoxon rank-sum test. The experimental results, supported by statistical evidence, demonstrate the superiority of the new EMCQ-based image segmentation strategy in terms of attaining optimal threshold values with balanced exploitation and exploration operations.
|
|---|