Automated Color Model–Based Concrete Detection in Construction-Site Images by Using Machine Learning Algorithms

Abstract

Concrete structural component detection in color images is a key pre-process in various applications such as construction progress measurement, structural health monitoring, and three-dimensional as-built modeling. The goal of this research was to identify an automated color model–based concrete detection method that (by using a machine learning algorithm) can detect concrete structural components in color images with a high level of accuracy. A data set consisting of more than 87 million pixels was generated from 108 images of concrete surfaces with a variety of surfaces. Transformations from the RGB color space to non-RGB color spaces were performed to increase separability between concrete and background classes and to achieve robustness to changes in illumination. To find the optimal combination of color space and machine learning algorithm, the performance of three machine learning algorithms (e.g., a Gaussian mixture model, an artificial neural network model, and a support vector machine model) in two non-RGB color spaces (e.g., HSI and normalized RGB) was comparatively analyzed. The comparison showed that the combination of the support vector machine algorithm and the HSI color space is superior in detecting concrete structural components in color images, compared with the other five algorithm–color space combinations. Performance was validated by experiments run on various images of actual construction-site scenes.