Page 48 - MSAM-4-3

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

Table 4. Parameter settings for the faster R‑CNN model
Parameter Value Description
(Training+Validation): Test 9:1 Split of training, validation, and test sets
Training: Validation 9:1 Split of training and validation data
input_shape (600, 600) Each input image is resized to 600×600
Backbone resnet50 Used for feature extraction
anchors_size (4, 16, 32) Sets anchor box sizes during training, generating anchor boxes at three different scales to
detect objects of varying sizes
Freeze_Epoch 50 End epoch for the frozen stage
Freeze_batch_size 4 Batch size during the frozen stage
Freeze_lr 1e-4 Learning rate during the frozen stage
nFreeze_Epoch 100 End epoch for the unfrozen stage
Unfreeze_batch_size 4 Batch size during the unfrozen stage
Unfreeze_lr 1e-5 Learning rate during the unfrozen stage
Confidence threshold 0.5 Predicted boxes with confidence scores above 0.5 are considered valid detections during
object detection
nms_iou threshold 0.3 Intersection Over Union threshold for non-maximum suppression; when the overlap of two
predicted boxes exceeds 0.3, the model retains the higher-scoring box and suppresses the
other to reduce redundancy
Test anchors_size (4, 8, 16) Anchor box sizes used during model evaluation
R-CNN: region-based-Convolutional neural network

faster inference speed, and sound detection performance. Figures 2 and 3, respectively. Both training and validation
YOLOv5 employs mosaic data augmentation, which losses for ResNet-50 decrease rapidly in the initial stages
combines four images into one input image, enhancing and stabilize, with closely aligned trends, indicating good
the detection of small objects. The model also provides an model fitting and effective learning without overfitting.
automated hyperparameter optimization system that searches However, its validation accuracy fluctuates, suggesting
for the best training configuration for hyperparameters, such weaker generalization. In addition, the training accuracy
as learning rate, weight decay, and mosaic probability. 39 of EfficientNetV2B0 surpasses 90% within the first few
The model training consists of two stages: Freezing batches and remains stable at around 100%, demonstrating
and unfreezing. First, the dataset and pre-trained weights superior learning capability and stability compared to
are loaded (to accelerate training and enhance feature ResNet-50. To ensure a fair comparison, both models were
extraction capabilities). In the freezing stage, only the trained for 50 epochs, but EfficientNetV2B0 exhibited slight
detection head parameters are trained, while the backbone overfitting in the final stages, as indicated by declining
network weights remain unchanged to reduce memory training loss, increasing validation loss, and fluctuations in
usage. In the unfreezing stage, the backbone network validation accuracy.
parameters are unlocked, allowing the entire model to The training and validation loss changes during the Faster
participate in training and improve performance. During R-CNN training process are displayed in Figure 4. In the first 50
training, loss values are recorded, and model weights epochs of the freezing phase, the training and validation losses
are saved after each epoch. A training log is generated to rapidly decrease in the initial few epochs and then gradually
evaluate the model’s performance. stabilize. The smoothness of the training and validation
The main settings and parameters used for the losses, with a difference of <0.1, indicates that the model did
Faster R-CNN and YOLOv5 models are presented in not experience significant overfitting or underfitting. During
Tables 4 and 5, respectively. the following 50 epochs, when the unfreezing phase begins
and the model undergoes full parameter updates, fluctuations
3. Results in the loss are normal. However, after that, the loss decreases

3.1. Loss and accuracy but does not fully converge.
The loss and accuracy changes for the image classification One limitation of the Faster R-CNN model is that it
models, ResNet50 and EfficientNetV2B0, are featured in selects 256 mini-batches of anchor boxes from the same

Volume 4 Issue 3 (2025) 6 doi: 10.36922/MSAM025150022