Classification of Building Cracks Image Using the Convolutional Neural Network Method

Abstract

Building crack images classification caused by the earthquake is commonly conducted manually by analyzing walls, beams, columns, and floors based on visual inspection of crack's diameter, depth, and length. Experts in structural engineering who have many experiences in building damage assessment usually handle the mentioned task. In order to speed up and simplify the assessment process a classification system based on pattern recognition is on demand. This paper proposes a crack image classification technique using CNN. This classification technique is proposed to improve the performance of two previous works: the crack classification systems using GLCM features and the SVM classifier and the crack classification systems using Zoning and Moment features and QDA classifier. The experimental results show that the CNN based crack image classification works properly indicated by 99.63% of accuracy, 99.65% of precision, and 99.64% of recall for METU dataset and 93.80% of accuracy, 93.49% of precision, and 93.94% of recall for CDLE dataset. In detail, the CNN based crack image classification provides significantly higher performance than that of the previous works. Furthermore, the proposed system also shows robust performance against large variability of cracks and non-crack images.