Automatic detection of concrete cracks from images using Adam-SqueezeNet deep learning model

Abstract

Cracks on concrete surface are typically clear warning signs of a potential threat to the integrity and serviceability of structure. The techniques based on image processing can effectively detect the cracks from images. These techniques, however, are generally susceptible to user-driven heuristic thresholds and extraneous distractors. Inspired by recent success of artificial intelligence, a deep learning based automated crack detection system called CrackSN was developed. An image dataset of concrete surface is collected by smartphone and carefully prepared in order to develop and train the CrackSN system. This proposed deep learning model, built on the Adam-SqueezeNet architecture, automatically learns the discriminative feature directly from the labeled and augmented patches. Hyperparameters of SqueezeNet are tuned with Adam optimization additive through the training and validation procedures. The fine-tuned CrackSN model outperforms state-of-the-art models in recent literature by correctly classifying 97.3% of the cracked patches in the image dataset. The success of CrackSN model demonstrated with light network design and outstanding performance provides a key step toward automated damage inspection and health evaluation for infrastructure.