Page 63 - Read Online

Hu et al. J. Mater. Inf. 2025, 5, 44 https://dx.doi.org/10.20517/jmi.2025.21 Page 13 of 17

mAP improvement when freezing the first layer compared to full-model fine-tuning (frozen_stages = 0).
However, the freezing strategy requires dynamic adjustment based on data scale and material diversity.

This study investigated the impact of varying pre-trained epochs during the self-supervised pre-training
phase on object detection performance on the NEU-DET dataset. As shown in Figure 7, there is a
significant correlation between pre-trained epochs and detection accuracy for both the NEU and SSDD
datasets. The mAP initially rises but then plateaus and slightly declines after reaching an optimal number of
epochs. This implies that increasing pre-training epochs beyond a certain point does not significantly
enhance detection capability and may even cause slight performance degradation. These findings highlight
the importance of selecting an appropriate number of pre-training epochs to optimize model performance.
It is worth noting that the NEU dataset consists of 1,800 images, while the SSDD dataset contains 20,272
images. This difference in dataset sizes explains why the NEU dataset required pre-training epochs ranging
from 100e to 800e, whereas the SSDD dataset only required epochs from 100e to 400e. The larger SSDD
dataset provides more diverse samples for the model to learn from, thus requiring relatively fewer epochs to
achieve satisfactory performance. In contrast, the smaller NEU dataset necessitates more epochs to ensure
the model can adequately learn the features needed for effective defect detection.

For the NEU dataset, as the number of pre-training epochs increased from 100e to 400e, both mAP and
mAP_50 showed an upward trend, indicating that the model gradually learned more effective features from
the NEU dataset, thereby improving its ability to detect steel surface defects. However, after a certain
number of epochs, the increase in mAP and mAP_50 slowed down and began to stabilize, suggesting that
the model’s performance may reach a saturation point with further pre-training epochs. Regarding the
SSDD dataset, the model achieved the highest mAP of 0.3850 and mAP_50 of 0.7630 at 200e. When the pre-
training epochs continued to increase to 300e and 400e, both mAP and mAP_50 exhibited a downward
trend. This might be due to overfitting to the SSDD dataset as the number of pre-training epochs increased
beyond the optimal range, reducing the model’s generalization ability for steel surface defect detection. As
for the SSDD + NEU dataset, the model demonstrated relatively stable performance across different pre-
training epochs, maintaining mAP values above 0.3780 and mAP_50 values above 0.7420. This indicates
that combining the SSDD and NEU datasets for pre-training helps enhance the model’s robustness and
generalization ability. The SSDD + NEU-100e configuration achieved a mAP of 0.3810 and a mAP_50 of
0.7540, while the SSDD + NEU-200e configuration obtained a mAP of 0.3810. Although the performance
was slightly lower than that of the SSDD-200e configuration, it still outperformed the baseline model using
ImageNet pre-trained ResNet18 weights (mAP of 0.3800). This suggests that the fusion of multiple datasets
can provide the model with more diverse features, enabling better adaptation to the complexity of steel
surface defect detection tasks.

Overall, the choice of pre-training dataset and the number of pre-training epochs significantly influence the
model’s performance. The SSDD-200e configuration achieved the highest mAP value, while the SSDD +
NEU configuration exhibited stable and relatively superior performance, highlighting the advantages of
combining multiple datasets for pre-training. These findings offer valuable insights for future research in
optimizing pre-training strategies to enhance the performance of self-supervised learning models in steel
surface defect detection.

[46]
[45]
Our method achieves mAP_50 = 0.7680 on NEU-DET, surpassing SSDD-Net by 4%, and YOLOv7
(mAP_50 = 0.7290). Transfer learning with our weights consistently outperforms random initialization,
confirming self-supervised efficacy for steel defect detection.