An Explainable Vision Transformer-Based Ensemble Framework for Automated Tire Wear Assessment

This research presents an automated tire wear assessment system integrating Vision Transformer (ViT) with Explainable AI to address the limitations of traditional manual inspections, which suffer from inconsistency, subjectivity, and high labor costs. The proposed ensemble framework combines Vision Transformer, Canny edge detection, and texture analysis with data-driven weighting optimized on validation splits and temperature scaling calibration. The model was trained and evaluated on two tire tread datasets: Tire Textures from Kaggle (1,028 images) and TyreNet from Mendeley Data (1,698 images) under a rigorous cross-source (LODO) protocol for binary classification (Normal/Cracked). The calibrated ensemble achieves 91.15% accuracy (macro-F1: 0.911) on fold $A$ and 76.68% accuracy (macro-F1: 0.758) on fold $B$, while standalone modules attain 92.10% (ViT), 77.00% (edge detection), and 83.30% (texture analysis), with attention maps aligning over 90 % with expert-identified defects. These findings demonstrate the system's ability to provide accurate, interpretable, and actionable diagnostic reports for practical deployment in automotive safety applications.