Crack Detection Based on Generative Adversarial Networks and Deep Learning

This paper proposes a novel crack detection method using the three-stages detection model. Deep learning technology has been a focus of attention in the field of crack detection; however, it needs big data to train the corresponding network model. More training samples and the combination of multiple deep learning algorithms help to improve the detection performance. Therefore, this paper employed a generative adversarial network (GAN) model to generate abundant virtual crack images with similar features to real images, these virtual images are used to train the CNN classifier and DeepLab_v3+ respectively, and then the real images are used to evaluate the performance of the three-stages detection method. The results show that the proposed three-stages detection method has excellent detection effect on the crack detection is better than that of the control experiment (the NI_MIoU, NI_Accuracy, NI_F-score and NI_MCC are increased by 22.1%-55.6%, 5.2%-9.8%, 37.4%-40.0% and 6.2%-11.1% respectively)). These results demonstrate that the three-stages detection model has made a beneficial contribution to the crack detection.

Automated summary

This paper proposes a novel crack detection method using a three-stages detection model. By employing a generative adversarial network (GAN) model to generate virtual crack images, the method achieves excellent detection performance.