A Deep Convolutional Neural Network Framework for Automated Crack Detection and Classification in Concrete Bridge Structures

Abstract

Surface crack identification is critical for assessing structural health of concrete highway bridges, but is slow, subjective and requires skilled inspectors. In this study, we present a deep CNN (DCNN) model that can be used, directly, to detect the presence of cracks and classify them into clinically useful groups: longitudinal, transverse, and map/diagonal cracking, directly, from digital images of bridge decks, girders and piers. The model consists of four hierarchical convolutional blocks, each with batch normalization, rectified linear activation, and max-pooling, a global average pooling layer, and a fully connected classification head with dropout regularization. The model was trained and tested on a curated dataset of 24,000 labelled concrete surface images from field photographs and augmented by geometric and photometric transformations. The proposed network obtained an overall classification accuracy of 97.6% and a mean F1-score of 0.968 and an average inference time of 18 milliseconds per image on a single graphics processing unit, which outperformed three widely used baseline architectures (a shallow custom CNN, VGG16 and ResNet18) under the same conditions. The interpretability of the predictions was verified by the Grad-CAM visualization, which showed that the network focused on the crack regions, but not the background textured parts or lighting artifacts. The results show that the proposed framework provides a feasible, accurate, and low-cost solution to automated, drone/robot-assisted bridge inspection programs that can save inspection time, cost, subjectivity, and provide consistency in condition assessments.