A Statistically Grounded and Physics-Aware Vision Framework for Detecting Barely Visible Impact Damage (BVID) in...

Barely Visible Impact Damage (BVID) in heterogeneous polymer-matrix composites remains difficult to detect because subtle damage signatures are often masked by complex architectures, hybrid textures, and overlapping failure morphologies. This study therefore presents an experimentally grounded, physics-aware, and statistically validated vision-based inspection framework rather than a purely detector-centered benchmarking exercise. Real post-impact images were obtained from controlled low-velocity impact experiments on 20 composite architectures and 60 physical specimens, yielding approximately 2000 images across laminated, hybrid, textile-reinforced, and sandwich structures. The dataset was organized using a specimen-disjoint splitting protocol to prevent leakage across training, validation, and test subsets. To improve robustness while preserving physical realism, a physically grounded Albumentations strategy was developed using only physically admissible transformations and explicit exclusion of non-physical operations that could distort damage morphology or surface continuity. Model development was further complemented by a hybrid hardware workflow in which cloud-based GPU training was combined with deployment-oriented inference profiling on resource-constrained edge-like hardware, thereby linking detection accuracy to practical industrial feasibility. In addition, model performance was evaluated under a standardized training budget and validated through repeated runs, Friedman significance testing, and Holm-corrected Wilcoxon signed-rank pairwise comparisons to ensure error-controlled interpretation of inter-model differences. Across the evaluated compact YOLO families, YOLO26s delivered the strongest overall performance, reaching 0.841 mAP@0.5, 0.586 ± 0.004 mAP@0.5:0.95, and an F1-score of 0.809, while YOLO11s achieved the highest precision and YOLO26n remained competitive in recall with nano-level compactness. Overall, the results show that experimentally generated heterogeneous composite data, morphology-preserving augmentation strategy development, leakage-aware dataset design, deployment-oriented computational profiling, and statistically grounded validation together provide a more robust and application-relevant basis for automated BVID detection in polymer-matrix composite structures.