A comparison between computer vision- and deep learning-based models for automated concrete crack detection

Abstract

Systems subjected to continuous operation are exposed to different failure mechanisms such as fatigue, corrosion, and temperature-related defects, which makes inspection and monitoring their health paramount to prevent a system suffering from severe damage. However, visual inspection strongly depends on a human being’s experience, and so its accuracy is influenced by the physical and cognitive state of the inspector. Particularly, civil infrastructures need to be periodically inspected. This is costly, time-consuming, labor-intensive, hazardous, and biased. Advances in Computer Vision (CV) techniques provide the means to develop automated, accurate, non-contact, and non-destructive inspection methods. Hence, this paper compares two different approaches to detecting cracks in images automatically. The first is based on a traditional CV technique, using texture analysis and machine learning methods (TA + ML-based), and the second is based on deep learning (DL), using Convolutional Neural Networks (CNN) models. We analyze both approaches, comparing several ML models and CNN architectures in a real crack database considering six distinct dataset sizes. The results showed that for small-sized datasets, for example, up to 100 images, the DL-based approach achieved a balanced accuracy (BA) of ∼74%, while the TA + ML-based approach obtained a BA > 95%. For larger datasets, the performances of both approaches present comparable results. For images classified as having crack(s), we also evaluate three metrics to measure the severity of a crack based on a segmented version of the original image, as an additional metric to trigger the appropriate maintenance response.