Page 62 - Read Online
P. 62
Page 12 of 17 Hu et al. J. Mater. Inf. 2025, 5, 44 https://dx.doi.org/10.20517/jmi.2025.21
Figure 8. NEU-DET detection performance with varied pre-trained weights vs. benchmarks. (A) Comparative evaluation of mAP for
object detection across NEU, SSDD, and SSDD + NEU pre-trained weights in conjunction with benchmark test results; (B) Comparative
evaluation of mAP_50 for object detection across NEU, SSDD, and SSDD + NEU pre-trained weights in conjunction with benchmark test
results. mAP: Mean average precision.
Figure 9. Visual examples of detected defects on NEU-DET. (A) Crazing; (B) Rolled-in_scale; (C) Inclusion; (D) Scratches; (E) Patches;
(F) Pitted_surface. Left column: ground truth annotations; Right column: model predictions with confidence scores.
representations from unlabeled image data. Experimental results demonstrate that employing pre-trained
weights from NEU and SSDD for defect detection on the NEU-DET dataset yields an approximately slight
