Page 50 - MSAM-4-3

Materials Science in Additive Manufacturing AI-driven defect detection in metal AM

Figure 5. Performance of the YOLOv5 model

Figure 6. Confusion matrix from the ResNet50 (left) and EfficientNetV2B0 (right) models

predictions, while metrics, including accuracy, precision, relatively easy to learn the small- to medium-sized image
and recall, are derived from the matrix to provide a datasets generated by 3D printers, effectively capturing the
comprehensive assessment (Figure 6). key features. This demonstrates good potential for future
development and suggests that these models could handle
In the 265 test image samples (Table 6), ResNet50 achieved more complex AM image datasets.
perfect classification, while EfficientNetV2B0 performed
equally well, with only one “good” image misclassified as For the EfficientNetV2B0 model with deeper networks,
“defects” and all other samples correctly classified. the defect classification task appears relatively simple,
reaching over 90% accuracy within five iterations and
After confirming that there was no information remaining at nearly 100% thereafter. If not overtrained,
leakage, the results (Table 6) demonstrated that the image it outperforms ResNet50 by learning faster and
classification model could always complete the classification demonstrating excellent stability.
task perfectly, with almost no missed or incorrect
detections. The choice of different loss functions made little 3.3. Object detection models comparison
difference since both achieved nearly 100% accuracy. On As displayed in Figure 7, Faster R-CNN achieved an
the test set, both models achieved fast detection, with the accuracy of 46.25%, which is measured by the area under
EfficientNetV2B0 model demonstrating a faster detection the precision and recall curves, reflecting the overall
speed. This indicates that these two models found it detection performance. A higher recall suggests that more

Volume 4 Issue 3 (2025) 8 doi: 10.36922/MSAM025150022