HSV Color Space Conversion and Deep Learning Model Performance Analysis for Efflorescence Segmentation in Damage Detection

Abstract Efflorescence is a common defect observed on the surfaces of concrete and masonry structures, usually appearing as deposits of white crystalline substances. Traditional detection methods for efflorescence rely primarily on visual inspection; however, such methods are limited by subjective judgment and require extensive manual labor, which is both time-consuming and costly. To address this problem, this paper proposes a noncontact, nondestructive testing (NDT) method that integrates computer vision techniques with deep learning models for the automatic detection of efflorescence on building facades and concrete surfaces. In this study, deep learning–based image segmentation techniques are introduced, and five commonly used deep learning models—U-Net, PSPNet, DeepLabV3, MMSegmentation, and ConnectNet‍—are compared in terms of their segmentation performance under different color spaces. Experiments were conducted using both RGB (red, green, blue) and HSV (hue, saturation, value) images. The results indicate that all models exhibit superior segmentation performance on HSV images compared with RGB images. In particular, MMSegmentation and ConnectNet achieved outstanding performance with efflorescence detection accuracies of 95.8% and 95.6%, respectively, in the HSV color space. Compared with the RGB color space, the HSV color space shows significant advantages in resisting lighting variations and improving detection accuracy. This study advances the body of knowledge in construction engineering and management by proposing a noncontact NDT framework that uses HSV preprocessing to improve deep learning–based efflorescence segmentation. A comparative evaluation of five semantic segmentation models provides a benchmark for automated efflorescence assessment under varying environmental conditions.