Page 59 - Read Online

Hu et al. J. Mater. Inf. 2025, 5, 44 https://dx.doi.org/10.20517/jmi.2025.21 Page 9 of 17

RESULTS AND DISCUSSION
Results
Benchmark test on NEU-DET
To evaluate the effectiveness of the self-supervised learning approach for steel surface defect detection, we
established baselines on the NEU-DET dataset (six defect categories; 1,800 total images: 1,080 training, 360
validation, 360 testing; Table 1) using Faster R-CNN with random weights and weights from an ImageNet
pre-trained ResNet18. These baselines provide comparative benchmarks for subsequent self-supervised
learning evaluations.

Faster R-CNN was trained for 24 epochs on the NEU-DET training split, and the final accuracy is shown in
Figure 5. The object detection accuracy began to stabilize after 18 epochs. After 24 epochs of training with
randomly initialized weights, the achieved mAP and mAP_50 values were 0.0880 and 0.2800, respectively.
Conversely, after 24 epochs with weights from the ImageNet pre-trained ResNet18 model, the
corresponding mAP and mAP_50 values were 0.3800 and 0.7730, respectively.

Detailed analysis of object detection on NEU-DET
“Frozen_stages” on object detection
We trained a contrastive representation model via self-supervision on two unlabeled datasets (NEU, SSDD)
until loss convergence. The resulting weights were transferred to steel defect detection, with fine-tuning and
validation performed on NEU-DET. During detection, varying frozen_stages (0 vs. 1) revealed superior
mAP at frozen_stages = 1 [Figure 6]. This optimal configuration was consequently adopted.

Number of pre-training epochs on object detection
To determine whether downstream detection accuracy correlates with pre-training epoch count, we
transferred different pre-trained models to the object detection task following pre-training on the NEU and
SSDD datasets. We evaluated the effect of different pre-training epochs on the object detection performance
using the NEU-DET dataset. Figure 7A depicts the mAP for object detection across 24 epochs with the use
of pre-training weights NEU-100e to NEU-800e. These results indicate that the accuracy of object detection
initially increased with the number of pre-training epochs, then experienced a slight decline, and ultimately
stabilized. As shown in Figure 7B, the six types of pre-trained models from SSDD, when transferred to the
object detection task, produced similar results, with the peak mAP reaching 0.3850, corresponding to the
SSDD-200e pre-trained model.

Construction of the pre-training dataset
This section analyzes how pre-training datasets influence defect detection performance on NEU-DET. We
transferred weights from three protocols - NEU, SSDD, and SSDD + NEU (trained across various epochs) -
to the detection task, recording mAP and mAP_50 in Table 2. On the SSDD dataset, the model attained its
peak mAP of 0.3850 and mAP_50 of 0.7680 at 200 epochs. Beyond this point, as pre-training progressed to
300 and 400 epochs, both metrics exhibited a decline. In contrast, when trained on the SSDD + NEU
dataset, the model’s performance remained consistently robust across varying pre-training durations,
sustaining mAP values above 0.3780 and mAP_50 scores above 0.7420. Notably, models initialized with
NEU-derived pre-trained weights consistently underperformed compared to those pre-trained on SSDD or
SSDD + NEU datasets.

Self-supervised learning viability for steel defect detection was assessed by transferring pre-trained weights
to NEU-DET and comparing against benchmarks. As shown in Figure 8A, the model initialized with
random weights exhibited the lowest performance, with a mAP of 0.0880. In contrast, SSDD-200e weights