Deep Learning Based Multi-Classifier System for Crack Segmentation in Civil Structure Images

In civil structures, cracks are one of the initial signs of structure deterioration. Therefore, cracks identification is an essential task for structures maintenance. In this framework, automatic inspection of structures is a suitable replacement to the manual approaches. These automatic methods are mainly based on computer vision techniques which has had a growing interest in last decades. Crack segmentation is similar to edge detection problems; thus, it could be solved by edge detection methods. In this paper, cracks in civil structures such as concretes and pavements are segmented by a Convolutional Neural Network (CNN) based multi-classifier system which applies pixel-wise segmentation on images of cracks. The dataset we have used made of 537 three channel images with manual annotation maps. The final results claims that the proposed method with F-score of 86.8, has good performance which is superior compared to holistically nested networks like HED and DeepCrack.