Cracks detection in images of concrete structures using deep neural networks

Abstract

This paper is inserted in the context of image analysis, aiming at the automatic extraction of complex information with high precision. This study aimed to evaluate the performance of convolutional neural networks in classifying concrete images into two classes: (a) non-cracked and (b) cracked. For this purpose, VGG16, VGG19, and ResNet50 deep networks were employed with transfer learning through fine-tuning. The networks were re-trained and tested using a database of 40,000 images. After training, the networks were tested, achieving an impressive accuracy between 99.27% and 99.78%. This high accuracy level inspires confidence in using these predictive models. To assess the robustness of the models, visual gradients of the networksʼ attention focal points on the images were generated, showing that the models focus on and capture aspects of the photos that truly characterize the cracks. Based on the results, it can be concluded that convolutional neural networks are effective in classification problems involving concrete and can be applied in accurate inspections to assist engineers with high reliability regarding the results.