Page 60 - Read Online

Page 10 of 17 Hu et al. J. Mater. Inf. 2025, 5, 44 https://dx.doi.org/10.20517/jmi.2025.21

Table 1. Details of the NEU-DET dataset
Split Images Annotations Categories
Train 1,080 items 2,488 items 6 items
Test 360 items 855 items 6 items
Val 360 items 846 items 6 items

Table 2. Performance metrics of the NEU-DET dataset: mAP and mAP_50
NEU SSDD SSDD + NEU
Performance metrics
100e 400e 100e 200e 300e 400e 100e 200e
mAP 0.3160 0.3700 0.3790 0.3850 0.3790 0.3780 0.3810 0.3810
mAP_50 0.6910 0.7550 0.7630 0.7680 0.7420 0.7550 0.7540 0.7520
The defect detection results corresponding to the SSDD-200e weight are the best, with the optimal values emphasized in bold. mAP: Mean
average precision.

Figure 5. Comparative analysis of mAP and mAP_50 for object detection on the NEU-DET dataset with random initialization and
ImageNet pre-trained ResNet18. mAP: Mean average precision.

achieved peak mAP (0.3850), exceeding ImageNet-pre-trained ResNet18 (0.3800). SSDD + NEU weights
showed strong performance with mAP of 0.3810 for both 100e and 200e.

Figure 8B compares mAP_50 across initialization methods. All self-supervised weights outperformed
random initialization, with SSDD-200e achieving peak mAP_50 (0.7680) - approaching ImageNet-pre-
trained performance (0.7730).

Detection visualization
In the experiment using NEU-400e pre-trained weights on the NEU-DET dataset for defect detection, the
results demonstrated the model’s effectiveness and reliability. Figure 9 presents six defect cases (crazing,