An analogy between various machine-learning techniques for detecting construction materials in digital images

Abstract

Digital images and video clips collected at construction jobsites are commonly used for extracting useful information. Exploring new applications for image processing techniques within the area of construction engineering and management is a steady growing field of research. One of the initial steps for various image processing applications is automatically detecting various types of construction materials on construction images. In this paper, the authors conducted a comparison study to evaluate the performance of different machine learning techniques for detection of three common categorists of building materials: Concrete, red brick, and OSB boards. The employed classifiers in this research are: Multilayer Perceptron (MLP), Radial Basis Function (RBF), and Support Vector Machine (SVM). To achieve this goal, the feature vectors extracted from image blocks are classified to perform a comparison between the efficiency of these methods for building material detection. The results indicate that for all three types of materials, SVM outperformed the other two techniques in terms of accurately detecting the material textures in images. The results also reveals that the common material detection algorithms perform very well in cases of detecting materials with distinct color and appearance (e.g., red brick); while their performance for detecting materials with color and texture variance (e.g., concrete) as well as materials containing similar color and appearance properties with other elements of the scene (e.g., ORB boards) might be less accurate.