CONCRETE CRACK IDENTIFICATION IN THE PRESENCE OF NONCRACK OBJECTS USING MACHINE LEARNING

Abstract

This study proposes a machine learning-based concrete crack identification strategy using digital image processing. The proposed approach is particularly designed to identify cracks when concrete surface images include crack-like noncrack objects that are difficult to separate from cracks by using conventional digital image processing algorithms. The machine-learning-based approach of this study initially (in the training stage) requires a set of various concrete surface images in which both cracks and crack-like noncracks are included. Each image is converted to a binary image, which helps determine crack candidate regions (CCR) and classify them to either cracks and noncracks manually. To extract the important features of cracks and noncracks effectively, the convolutional neural network (CNN) and the speeded-up robust features (SURF) are employed for each CCR. A crack classification model is finally constructed after applying clustering and classification processes to the regions. The performance of the proposed crack identification approach is experimentally validated.