Date of Graduation
12-2025
Document Type
Dissertation
Degree Name
Doctor of Philosophy in Engineering (PhD)
Degree Level
Graduate
Department
Industrial Engineering
Advisor/Mentor
Liao, Haitao
Committee Member
Braham, Andrew
Second Committee Member
Rossetti, Manuel
Third Committee Member
Zhang, Shengfan
Keywords
UAV-based inspection, Computer Vision, Anomaly Detection, Roof Defect Detection
Abstract
Building structural integrity depends on efficient anomaly detection methods that identify defects before significant damage occurs. Crack-induced roof leakages from material deterioration exemplify subtle anomalies that escape traditional inspection until extensive deterioration manifests. Computer vision integrated with unmanned aerial vehicles (UAVs) enables automated inspection, but detecting small-scale defects from aerial imagery faces challenges from background noise and environmental variations. This research advances automated vision-based roof inspection through three complementary approaches. First, we develop MAE-YOLOv8, a multi-source data fusion framework that incorporates an Attention-Enhanced module (AE-c2f) into the YOLOv8 backbone. This innovation enables accurate detection of subtle percolation features by focusing on relevant visual information while suppressing background noise through multi-stage feature fusion via convolution and concatenation operations. Second, we propose an Attention-Enhanced Deep Neural Network (AE-DNN) that integrates spatial and channel attention mechanisms to enhance feature extraction from UAV-captured Red-Green-Blue (RGB) images. The framework employs a Mixed Loss Function based on Focaler- Intersection over Union (IoU) to improve detection accuracy through optimized bounding box regression. Third, we introduce an adaptive weighted coverage planning framework that dynamically prioritizes structurally relevant roof regions through multi-cue visual analysis. The method integrates multi-scale keypoint detection, dark-stain and joint feature extraction, color consistency analysis, and periodic-pattern suppression into an inspection priority map normalized to a 0–1000 scale. A greedy region-merging algorithm transforms high-weighted regions into non-overlapping candidate inspection blocks for path optimization. Experimental validation demonstrates superior performance across all three approaches. MAE-YOLOv8 achieves mAP@0.5 of 0.754 with 85% keypoint reduction through quota-based sampling while maintaining spatial fairness. AE-DNN attains precision of 0.784 and mAP@0.5 of 0.742, successfully identifying percolated areas under varying environmental conditions. The adaptive path planning framework consolidates inspection regions into 46 non-overlapping blocks with 0.2229 average overlap ratio. Comparative evaluation of three path generation algorithms—Nearest Neighbor (NN) + 2-opt, Randomized Approximate Neighbor (RAN) + 2-opt, and Simulated Annealing (SA)—reveals that NN + 2-opt achieves optimal balance between path quality and computational efficiency, completing route generation in 28.1 seconds compared to 868 seconds for RAN + 2-opt and 3505 seconds for SA. The modular design decouples detection from path optimization, establishing a foundation for scalable autonomous UAV-based inspection systems. Future work will extend these capabilities through closed-loop adaptive replanning enabling real-time trajectory adjustment, temporal analysis for defect evolution tracking, and probabilistic path planning under environmental uncertainties.
Citation
Tan, M. (2025). Adaptive Vision-based Methods for Building Anomaly Detection Using Unmanned Aerial Vehicles and Machine Learning. Graduate Theses and Dissertations Retrieved from https://scholarworks.uark.edu/etd/6030
