Semantic segmentation for crack detection via generative knowledge distillation

Abstract

Recently, deep learning has garnered significant attention for its potential to detect damage in infrastructure. This approach requires a vast dataset for optimal performance; however, acquiring large-scale training data remains challenging. To overcome this challenge, this paper proposes a new technique for enhancing crack detection accuracy by synthesizing virtual crack images through generative algorithms. To this end, generative adversarial networks are used for generating new insights for crack images, and these insights are subsequently integrated into crack detection models using knowledge distillation. The proposed method obviates the need for additional crack images and enriches the diversity of the dataset. This approach yields a 5.09% crack intersection over union and a 3.51% improvement in the F1-score across 17 neural network models, outperforming traditional supervised learning methods. The proposed method is expected to gain widespread adoption in the future to address data scarcity challenges and enhance the safety of infrastructure maintenance.