A deep residual neural network framework with transfer learning for concrete dams patch-level crack classification and weakly-supervised localization. (January 2022)

• Record Type: Journal Article
• Authors: Li, Yangtao; Bao, Tengfei; Xu, Bo; Shu, Xiaosong; Zhou, Yuhang; Du, Ye; Wang, Ruijie; Zhang, Kang
• Is Part Of: Measurement. Volume 188(2022)
• Publication Date: 2022-01
• Subjects: Structural health monitoring -- Automated inspection -- Image-based -- Transfer learning -- Crack classification -- ResNet; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2021.110641
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: A structural crack ImageNet from various civil infrastructures is proposed. A novel strategy containing both in-domain and cross-domain transfer learning is proposed. More convolutional layers are unfrozen for fine-tuning to capture feature maps of structural damages. It can classify and weakly-supervised locate the dam crack areas with high accuracy and recall. Abstract: Under the continuous cycle of the external and environmental loads, the material property of dam structures will inevitably degrade and various structural defects will occur frequently. Among them, crack is one of the most threatening structural diseases, which has aroused extensive research attention. Identifying these structural cracks with high efficiency and accuracy is of great significance to the reinforcement of dams and maintenance of national infrastructure investment. This paper proposes an automatic classification and weakly-supervised localization paradigm for structural cracks in concrete dams based on deep residual neural networks and transfer learning. Inspired by ImageNet Challenge, this paper firstly proposes a structural crack ImageNet containing cracks collected from various types of civil infrastructures. Considering the difference between structural crack images and natural images, a combined two-stage transfer learing strategy is proposed. Specifically, more convolutional layers of the pre-trained model on the ImageNet are unfrozen and fine-tuned on the proposed structural ImageNet. To improve the model interpretability, Gradient-weighted class activation mapping(Grad-CAM) approach is utilized to classify and weakly-supervised locate the underlying surface cracks on an arch dam without relying on refined manual data annotations. The experimental results demonstrate that the proposed framework can effectively identify structural cracks of concrete dams with high accuracy and recall without relying on refined manual annotations under strong environmental background inference conditions. The proposed framework can be equipped with image acquisition equipment to realize real-time detection and localization of concrete dam structural defects.