Structural Crack Detection and Classification using Deep Convolutional Neural Network

Abstract

Cracks are indicators that affect the stability and integrity of infrastructures. Fast, reliable, and cost-effective crack detection methods are required to overcome the shortcomings of traditional approaches. This paper works on a transfer learning approach based on the deep convolutional neural network model VGG19 to detect cracks. Further, the proposed method is based on an improved VGG-19 model. The experiment is carried out on the SDNET2018 annotated images dataset. The dataset comprises of total 15k images, training set consists of 5000 cracked and 5000 un-cracked images of walls, pavements, and bridges. The experimental results on the proposed model provide 91.8% accuracy in detecting cracks on the testing set. The paper concluded that fine-tuning of the VGG19 (Visual Geometry Group) model accomplish satisfactory results in detecting cracks on images of multiple infrastructures.