<![CDATA[Automated visual inspection increasingly relies on deep learning to identify machine-component damage with high accuracy and consistent performance. However, direct comparison among architectures remains difficult because differences in data preparation, training procedures, and evaluation protocols can substantially affect reported results. This study quantitatively evaluates four deep learning architectures—conventional convolutional neural network, ResNet50, EfficientNet-B0, and MobileNetV3—for multi-class classification of machine-component images comprising Normal, Crack, Wear, Deformation, and Corrosion categories. A dataset of 1,230 images was processed using standardized preprocessing, stratified 70:15:15 partitioning, training-only augmentation, and a common model-development procedure. Performance was assessed using accuracy, macro-precision, macro-recall, macro-F1-score, confusion matrices, repeated five-fold cross-validation, feature-importance analysis, and computational indicators. On the independent test set, the baseline convolutional neural network, EfficientNet-B0, and MobileNetV3 achieved perfect classification performance, with accuracy and macro-F1-score of 1.000, whereas ResNet50 achieved 0.989 accuracy and 0.988 macro-F1-score. Cross-validation provided a more nuanced result: ResNet50 obtained the highest mean macro-F1-score (0.9934) and the lowest variability (SD = 0.0021), while the difference between the baseline convolutional neural network and EfficientNet-B0 was not statistically significant. No model misclassified a damaged component as Normal. Feature-importance analysis identified a concentrated contribution from selected image-derived descriptors, while MobileNetV3 demonstrated the lowest computational burden. The findings demonstrate that predictive performance, stability, interpretability, and computational efficiency should be jointly considered when evaluating deep learning models for industrial visual inspection.]]>
Copyrights © 2026
