Complete end-to-end pavement distress assessment using low-cost embedded systems with StyleGAN3 synthetic data augmentation for budget-limited transportation agencies

Traditional pavement monitoring systems costing hundreds of thousands of dollars create significant barriers for budget-limited transportation agencies. This research addresses this accessibility gap by developing a complete computer vision framework for automated pavement distress detection and quantification using low-cost embedded systems ($214) combined for affordable data collection with synthetic data augmentation to overcome fundamental class imbalance challenges. The framework employs StyleGAN3-based synthetic data generation with automated quality assessment (94.12% accuracy) and evaluates three integration strategies: direct synthetic augmentation, iterative pseudo-label refinement, and teacher-student transfer learning. Systematic evaluation across eight state-of-the-art architectures (YOLOv9-12, RT-DETRv2, Mask R-CNN variants, RTMDet) reveals that optimal synthetic data ratios vary substantially by architectural design, ranging from 60% to 300% of original dataset size, with transformer-based models showing earlier saturation compared to CNN-based architectures. The methodology integrates advanced image stitching for georeferenced orthomosaic creation and automated distress quantification. Comprehensive validation across twelve road segments (7.15 miles) demonstrated substantial performance improvements up to 12.3 percentage points in mAP@0.5:0.95, with direct augmentation proving most effective (0.896 mAP@0.5 detection, 0.905 segmentation). A complete deployment package including CPU-based graphical user interface and GIS integration enables practical implementation for transportation agencies, providing accessible state-of-the-art pavement monitoring capabilities without requiring specialized computational infrastructure or technical expertise.