Deep learning-based surface and subsurface damage identification using computer vision and thermography

The early detection of both surface and subsurface damage is crucial for ensuring structural integrity. Timely repair of damage also delays the need for infrastructure replacement, which involves significant costs and has adverse environmental impacts. While manual inspection for damage detection is a common practice, it is expensive, time-consuming, and hazardous. Moreover, it cannot easily cover all structures. To enable safe and autonomous detection of surface and subsurface damage, an automated Structural Health Monitoring (SHM) system is required. In this thesis, deep learning-based methods are proposed for detecting external and internal damage in structures using computer vision and active thermography, respectively. Additionally, an automated SHM system was developed by integrating these deep learning-based SHM methods with autonomous flight capabilities of unmanned aerial vehicles (UAVs). For internal damage, a new internal damage segmentation network (IDSNet) was employed for pixel-wise subsurface damage segmentation. IDSNet comprises advanced deep learning operators such as the intensive module, residual intensive convolution module, and superficial module. These operators enable IDSNet to process large thermal images in real-time with high accuracy, reducing monitoring costs. To overcome the challenges of costly and time-consuming ground truth data collection, an attention-based generative adversarial network (AGAN) was developed to generate synthetic image data for training IDSNet. The IDSNet demonstrates superior performance compared to other networks in accurately segmenting internal damages using active thermography. In addition to subsurface damage, surface damage, such as pavement potholes, is of significant concern. This thesis introduces 3DPredicNet, a novel monocular deep learning-based method for pothole segmentation with 3D volume prediction. The 3DPredicNet incorporates an advanced attention mechanism to reduce the number of learnable parameters. A dataset was prepared to train and test the model, and its robustness was also evaluated using a publicly available dataset. Lastly, this thesis integrates a computer vision-based damage detection method with an autonomous UAV system capable of navigating in GPS-denied areas. The proposed approach focuses on real-time multiple-surface damage detection using an improved faster region-based deep convolutional neural network and autonomous UAV. Specifically designed for GPS-denied structures, it mitigates the risks faced by inspectors during data collection in remote areas.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.