Crack Detection and Location Estimation Using a Convolutional Neural Network

Mengling Ding, Kwanghee Won

Abstract: In this paper, we present a CNN-based crack detection approach. Instead of classifying each pixel or an entire image patch, the proposed approach evaluates 48 predefined image stripes (in horizontal, vertical, and diagonal directions), and determine whether each area contains a part of crack or not. In the experiment, the CNN trained with about 1200 labeled images showed 75.9% of detection accuracy.