Utilizing pretrained convolutional neural networks for crack detection and geometric feature recognition in concrete surface images

Accurate detection of concrete cracks and identification of crack geometric characteristics are essential for ensuring structural safety. To address this, a concrete surface crack detection method was developed using pretrained convolutional neural networks (CNNs). Additionally, a comprehensive framework was proposed that integrates the pretrained CNN as a feature extractor with various regression algorithms to recognize concrete crack features, including the crack area, maximum width, and average width. Results demonstrate that the detection accuracy and training speed of modified CNN models based on pretrained networks, such as VGG16 and MobileNet, outperform those of CNN models trained from scratch. Moreover, the established CNNs achieve high accuracy in handling diverse images affected by environmental disturbances and noise. The developed comprehensive framework successfully recognized crack geometric features in concrete surface images. Evaluation of four regression algorithms revealed that support vector regression (SVR) achieved R2 values of 0.863 for predicting crack area and 0.764 for predicting the crack average width, while the XGBoost regression algorithm yielded an R2 of 0.782 for predicting the crack maximum width.