CrackNet: Pavement Crack Detection and Classification Based on Deep Learning Models

The identification of pavement cracks is essential for reducing traffic accidents and minimizing road maintenance costs. Existing crack detection methods frequently encounter challenges related to inefficiencies and accuracy, resulting in billions of dollars spent globally on road repairs annually. This study introduces an enhanced deep learning-based network aims at improving the accuracy of pavement crack detection. The proposed CrackNet utilizes advanced geometric augmentation techniques to enhance the model performance in identifying cracks across a variety of road conditions. We introduced CrackNet, a custom-designed deep-learning classification framework for pavement images. CrackNet combines pre-trained backbone feature extractors (MobileNetV2, InceptionV4, SqueezeNet) with a lightweight classifier head and a comprehensive preprocessing + augmentation pipeline to improve generalizability and address class imbalance. We evaluated three CrackNet variants (Sqz-CrackNet, Mob-CrackNet, Incep-CrackNet), each distinguished by its backbone, and found Incep-CrackNet achieved the highest accuracy of 96.05%, surpassing the Mob-crackNet and Sqz-crackNet, which attained accuracies of 95.44% and 90.06%, respectively. These findings underscore the effectiveness of the proposed deep learning framework in accurately detecting pavement cracks, representing a significant advancement over traditional detection methods.