Page 17 - Read Online

Page 10 of 28 Zhang et al. J Mater Inf 2024;4:34 https://dx.doi.org/10.20517/jmi.2024.64

indicated in Figure 6, the data augmentation methods encompass image flipping, random rotation, resizing,
cropping and adjustment to lightness, saturation, contrast and color. Additionally, CNNs are heavily
dependent on training samples, and changes in the target task or application routinely necessitate retraining
and re-annotation, potentially reducing development efficiency. To address the problems above, the transfer
learning method is introduced, which applies knowledge learned from one task to another related task [89,90] .
The models are pre-trained with the ImageNet dataset, which contains millions of labeled images
commonly implemented for image recognition tasks . The pre-training enables the model to capture
[91]
general image features and significantly accelerates the entire training efficiency.

Two strategies are typically adopted for updating weights in transfer learning models: partial layer freezing
and full fine-tuning. Partial layer freezing fixes some layers after loading pre-trained weights, allowing only
the remaining layers to be trained. Full fine-tuning utilizes pre-trained weights as initialization and updates
the entire model with a lower learning rate, leveraging prior knowledge to speed up convergence.
Determining the optimal freezing range necessitates extensive experimentation, which can be
time-consuming and resource-intensive. Therefore, pre-trained weights are initially loaded during the
construction of the defect recognition model, followed by comprehensive fine-tuning. Training parameters
are shown in Table 3. Due to the memory limitations of the RTX 3050 GPU, eight images per batch are used
throughout local training, while 32 images per batch are used on the NVIDIA A100 platform.
Supplementary Materials include the programs employed for the training and validation of CNNs in the
research.

Performance evaluation of neural network models
Incomplete fusion of the weld seam during pulsed TIG welding can result in flaws that undermine the
welding strength and overall quality. The CNN architectures including ResNet34, MobileNetV2, and
ConvNeXt are placed into experiment to examine the unfused defect on pulsed TIG welding images. Two
ResNet34 training experiments are conducted: one with randomly initialized weights, and the other with
weights pre-trained on the ImageNet dataset. Both models are then trained on welding images from pulsed
TIG experiments. As shown in Figure 7, both versions of ResNet34 initially had a validation accuracy of
around 0.55, indicating no classification capability. However, the classification accuracy of pre-trained
ResNet34 swiftly ascends, reaching 93.41% after 500 iterations. In contrast, the accuracy of randomly
initialized ResNet34 grows sluggishly, plateauing at 78.18%. Although further improvement is possible, it
would require substantial computational resources, making it less cost-effective for defect classification.
Furthermore, it is worth noting that the improvement of pre-trained ResNet34 significantly diminishes after
the initial surge. The accuracy has increased marginally from 93.31% to 93.79% until 5000 epochs, indicating
that the model’s performance remained at a stable level after the initial rapid rise and was difficult to
continue to improve.

The confusion matrices provide a detailed evaluation of model performance, presenting precision and recall
values for CNNs as shown in Table 4. In the research, the fully fused state is defined as positive, while the
unfused state is regarded as negative. Precision represents the proportion of true positives among all
predicted positives; high precision ensures that samples predicted as fused are indeed fused. Recall measures
the ratio of identified positive samples to all actual positive samples; high recall guarantees that no fused
samples are overlooked by the classification model. In welding defect recognition, all positive samples are
directly utilized in the workflow, making it essential to ensure that these samples are genuine positives.
Therefore, precision takes precedence over recall. As shown in Figure 8, ResNet34 achieved a precision of
96.52% and a recall of 92.5%, meeting industry requirements effectively. The defect classification model
tends to classify well-fused images as non-fused, a trend that supports timely control measures during the