DeepInspect: a deep learning tool for automated structural health monitoring with crack analysis

Purpose The evolution of computer vision algorithms has enabled us to replace redundant tasks, such as continuous monitoring with simple camera installations. This paper proposes a lightweight crack detection mechanism for deployment on mobile phones to automate the structural health monitoring (SHM) task in construction. Traditional sensor-based and human expert-guided monitoring tools are costly and subjective in nature. Thus, automatic detection increases performance with substantially fewer resources. Design/methodology/approach To provide a stable solution for crack detection, the authors work with two pipelines: one for identifying cracks from images and the second for identifying cracks directly from a mobile feed, and propose a novel ensemble of MobileNetv2 and UNet for crack detection in images and use YOLOv5n to detect cracks in a live stream. For training and testing, the authors create a comprehensive dataset of 9K images with different annotation schemes, where 5,600 images were annotated for image classification, 1,200 image masks were created for segmentation, and 2,200 boundary box annotations were made for object detection. Findings The proposed model achieved 98% accuracy for the MobileNetv2 module; 97% accuracy for the UNet module, and a [email protected] of 79.9 was achieved on the YOLOv5n module of the proposed data set after training. The model is then converted to a lightweight TFLite model and deployed on Android with the proposed Android application. The code and data are available at Link to the cited article. Research limitations/implications The deviation in angle leads to error susceptibility in crack severity. The effect of weather on images for crack detection is very limited. The cross-training of the surfaces also results in low crack detection accuracy by the YOLO model. Practical implications This study is useful for the automated continuous image-based structural health management for critical structures. Originality/value This work provides a practical and real-time solution to the complex problem of crack detection and SHM. Including advanced vision models with lightweight Android-based applications, the solution benefits site engineers and surveyors.