Deep learning methods have significant potential in facilitating fully automated pavement crack detection for highway maintenance projects. One of the biggest challenges to adopting these methods is the lack of a high-quality balanced training datasets (i.e. images of labelled cracks). Data imbalance is one of the biggest causes of detection bias and overfitting in trained models. The reason maintaining a balanced training dataset is challenging in case of crack detection tasks specifically, is because different types of cracks occur in an inherently unbalanced manner (i.e. some cracks are more common than others). Also, the variety of cracks available in different areas can be limited. This results in overfitting of models to a certain type of cracks and their failure to detect other uncommon cracks. Manual collection of such a specific dataset can be expensive, time consuming, and labour intensive. To solve data scarcity in real-world applications, this research employs novel generative adversarial networks (GANs) to produce life-like artificial crack data that is then used to augment existing training datasets and improve the classification accuracy. The images in benchmark Crack500 and CrackForest datasets were first annotated and segregated into four classes, transverse, longitudinal, and alligator cracks, to create a baseline for our models. Data augmentation was then implemented by adding artificial crack image data generated using GAN. We then examined how artificial data generated by each network improved the classification model’s accuracy when compared to the baseline where no artificial data was added. We demonstrate that GAN models can generate a variety of quality life-like crack images that can improve the overall classification results and metrics significantly. The proposed approach addresses the limitations of imbalanced datasets, which enhances our ability to automatically detect multiple pavement crack types using AI technology.