Automatic recognition of concrete spall using image processing and metaheuristic optimized LogitBoost classification tree

This paper presents a novel artificial intelligence model to automatically recognize concrete spall appearing on building components. The model is constructed by integrating a metaheuristic optimization algorithm, advanced image processing techniques, and a powerful machine learning-based inference model. Kapur's entropy based image segmentation, statistical measurements of image color, gray level co-occurrence matrices, and local ternary pattern are used to extract numerical features presenting concrete surfaces on spall and non-spall samples. Subsequently, a LogitBoost based ensemble framework of classification and regression tree (CART) models (denoted as LBT) is employed to construct a decision boundary capable of recognizing spall/non-spall image samples. Moreover, in order to enhance the performance of the LogitBoost based ensemble framework, forensic-based investigation (FBI) metaheuristic is utilized to determine the most suitable set of the framework's hyper-parameters including the learning rate (μ), the learning cycle (L_c), the minimum number of leaves (L_min), and the maximum number of splits (S_max). A data set including 486 image samples has been collected from field surveys at high-rise buildings in Da Nang city (Vietnam) to train and verify the proposed FBI optimized LBT model (denoted as F-LBT). Experimental results supported by statistical tests point out that the F-LBT is a capable method for concrete spall detection with a classification accuracy rate = 88.3%, precision = 0.889, recall = 0.874, F1 score = 0.881, and negative predictive value = 0.874. Hence, the proposed hybrid approach is a promising tool to support building maintenance agencies in the task of periodic structural inspection.