This study compares CNN architectures for predicting the composition of WC–Co cemented carbides from SEM microstructure images. Because microstructural inhomogeneity formed during powder sintering can degrade mechanical performance, image-based comp...
This study compares CNN architectures for predicting the composition of WC–Co cemented carbides from SEM microstructure images. Because microstructural inhomogeneity formed during powder sintering can degrade mechanical performance, image-based composition assessment is important for quality and process control. Microstructure images of WC–Co with different Co contents were collected and converted into a trainable dataset through up-scaling, sliding-window cropping, rotation, and resizing. EfficientNet, ResNet, and MobileNet were applied to this dataset, and their classification performance was analyzed using accuracy, precision, recall, specificity, F1-score, and confusion matrices.
The results indicate that CNN-based classifiers can accurately predict WC–Co composition from SEM images, with EfficientNet achieving the most balanced overall metrics and confirming the feasibility of image-based, data-driven composition assessment.