Deep Learning-Based Quality Classification of Cold Spray Coatings with Explainable Feature Analysis

This study investigates the applicability of deep learning models for automated quality classification of cold spray coatings, focusing on three deposition categories: good, degraded, and poor deposition. Three state-of-the-art convolutional architectures, ResNet-50, EfficientNet-B0, and ConvNeXt-Tiny, were evaluated across two training phases designed to assess the impact of dataset balancing, data augmentation, and higher input resolution. In the first phase, models were trained on an imbalanced dataset using only class weighting; EfficientNet-B0 achieved the best performance (ACC 80%, F1 77%), while ResNet-50 showed notable instability (ACC 60%, F1 56%). In the refined second phase, oversampling, advanced augmentation, 380×380 resolution, and early stopping led to substantial performance gains for all models. ConvNeXt-Tiny achieved the most robust and balanced results (ACC 93.3%, F1 90.3%), outperforming EfficientNet-B0 and ResNet-50 particularly in sensitivity and specificity for minority classes. Grad-CAM analysis provided qualitative insights into the decision-making process: poor samples elicited strong, spatially extended activations corresponding to defective regions, degraded samples produced more localized responses aligned with mid-scale irregularities, and good samples yielded diffuse, low-intensity activation patterns associated with surface uniformity. These interpretable attention maps validated the physical relevance of the learned features and confirmed the suitability of ConvNeXt-Tiny for reliable and explainable cold spray quality assessment.