Enhancing crack segmentation accuracy in varied conditions using least-squares GAN and Pix2Pix

The detection and assessment of damage in civil infrastructure, particularly crack formation, are critical for early damage identification and effective maintenance planning. This paper investigates crack segmentation under varying environmental conditions using Pix2Pix-CrackNet, a deep learning-based neural network. The Pix2Pix-CrackNet model was trained on 6903 image pairs and validated using 1695 image pairs. Several strategies were implemented to improve crack segmentation accuracy, including the integration of features from the least squares generative adversarial network (LSGAN) and Pix2Pix, as well as optimisation of discriminator training relative to generator training through the use of a threshold value. To validate the performance of Pix2Pix-CrackNet, two established segmentation networks, namely U-Net and U-Net++, were also trained and evaluated using the same dataset. The performance of these models was then compared with that of Pix2Pix-CrackNet. Based on the evaluation of the segmentation models using Intersection over Union (IoU), precision, recall and F1-score metrics, Pix2Pix-CrackNet demonstrated superior performance compared with U-Net and U-Net++ on the test dataset. It achieved IoU, precision, recall and F1-score values of 72%, 87%, 53% and 66%, respectively, confirming its effectiveness and stronger generalisation capability for crack segmentation.