Color image segmentation based multilevel thresholding for detection of cracks and bug-holes in concrete surface images using optimization methods

Abstract

Automation in crack and bug-hole detection through non-destructive evaluation methods remains a major challenge in Structural Health Monitoring. Typically, machine learning models are fed with original, grayscale, or binary formats of concrete images. Original images retain excessive color data, grayscale formats reduce color without preserving structural relevance, and binary images often lack sufficient detail for assessing damage type and severity. To overcome these limitations, multilevel colored image thresholding using meta-heuristic algorithms such as Particle Swarm Optimization, Genetic Algorithm, Jaya and Sine Cosine Algorithm—guided by Otsu’s objective function. This technique enhances the contrast, edges, and texture boundaries, which serve as crucial primitives for object recognition, compared to original images. Additionally, it increases pixel connectivity, thereby simplifying image analysis for deep learning and machine learning applications. In fact, this method reduces the number of colors by an average of 97.3%, significantly decreasing computational load, while maintaining a high Structural Similarity Index (SSIM) of 0.873. Experimental results demonstrate that the optimized images outperform the original, grayscale, and binary formats in both object detection performance and precise evaluation of damage severity and progression. This fusion of optimization and perceptual image representation presents a promising advancement for automated structural damage assessment.