Automatic classification of asphalt pavement cracks using a novel integrated generative adversarial networks and improved VGG model

Crack development is increasingly intensified and causes pavement damage in recent decades under extreme weather events. Although various auto- or semi-auto crack classification algorithms have been proposed, most of them require manual extraction of image features, which is considerably labor-intensive, compromising classification accuracy and efficiency. Moreover, collecting original images for model training is difficult due to various limitations. This study proposes a Generative Adversarial Networks (GAN)-based method for data augmentation of the collected crack digital images and a modified deep learning network (i.e., VGG) for crack classification. Firstly, according to the characteristics of collected data, a GAN-based image generation model is established to expand the training dataset. Then, an improved VGG model is built based on the most efficient model via comparisons of several mainstream feature extraction networks. Finally, comparison studies of classification performance are performed for different classification models (i.e., the improved VGG and other traditionally used ones) and datasets (i.e., generated by GAN-based and traditional methods). The model trained by the dataset expanded by GAN has a higher accuracy rate and lower loss values than traditional methods. The improved VGG model trained by the validation set performs similarly to the training set. Compared to the original VGG model, the accuracy of crack prediction of the improved VGG model is increased by 5.9% (i.e., 96.30%), and the F1-score is also increased by 5.78% (i.e., 96.23%). Trained by the same test set expanded by GAN, the improved VGG model has a higher recall and F1-score than GoogLeNet, ResNet18, and AlexNet. The novel integrated GAN and modified VGG model shows satisfactory efficiency for classifying pavement cracks.