Page 64 - Read Online

Page 14 of 17 Hu et al. J. Mater. Inf. 2025, 5, 44 https://dx.doi.org/10.20517/jmi.2025.21

Self-supervised learning methods have demonstrated impressive accuracy on the defect detection task using
the NEU-DET dataset. To further demonstrate the superiority of the prediction accuracy of the SimSiam
model, we also present the prediction results of another self-supervised model (SimCLR) on this defect
detection task in Supplementary Figure 1. The results show that the mAP and mAP_50 prediction values of
the SimSiam model are consistently higher than those of SimCLR.

Cross-dataset transferability demonstrates that learned features generalize across metal types, capturing
universal surface characteristics [Supplementary Table 1]. Scratch and patch defect categories exhibit robust
detection performance, achieving peak mAP_50 values of 0.9780 and 0.9320, respectively, attributable to
their distinctive morphological characteristics. Conversely, crack defects demonstrate significantly
diminished accuracy with maximum mAP_50 of only 0.4720, likely resulting from subtle visual
manifestations or inadequate training sample representation. Regarding weight initialization strategies,
SSDD and SSDD + NEU pre-trained weights show superior domain-specific adaptation, consistently
outperforming both random initialization and NEU pre-trained configurations, while ImageNet-pre-trained
ResNet18 maintains competitive detection capability across most defect categories.

Limitations of self-supervised learning in object detection applications
This study reveals that self-supervised pre-trained weights transferred to downstream defect detection tasks
consistently underperform ImageNet-supervised ResNet18 baselines in mAP_50 metrics. While self-
supervised approaches achieve competitive accuracy, their inability to surpass large-scale supervised pre-
training underscores fundamental limitations in feature generalizability. We attribute this gap to two
interrelated factors: (1) constrained diversity in domain-specific unlabeled datasets (e.g., steel defect
imagery), which restricts comprehensive feature representation learning; and (2) inherent architectural
discontinuities between contrastive pre-training (e.g., SimSiam) and task-specific fine-tuning, inducing
feature subspace mismatches that undermine transfer efficacy.

Practical implementation faces additional challenges: Computational demands for industrial-scale datasets
necessitate prohibitive pre-training durations. Current data augmentation pipelines, though effective for
texture-based defects, fail to capture the full spectral and topological variance of real-world metallic
degradation phenomena.

Future research should focus on three cooperation pathways: hybrid semi-supervised frameworks that
leveraging limited labeled exemplars alongside abundant unlabeled data to bridge representation gaps; and
meta-transfer learning for few-shot adaptation to novel defect morphologies. These advances would
establish resource-efficient pipelines adaptable to the dynamic feature landscapes of industrial metal
inspection without compromising detection robustness.

CONCLUSIONS
Drawing inspiration from groundbreaking applications of self-supervised learning in scientific research,
such as AlphaFold2 - which leverages unlabeled data to solve complex problems - our study underscores the
potential of self-supervised methods as viable alternatives or supplements to traditional supervised learning
techniques in steel surface defect detection. The innovative use of unlabeled data significantly reduces
reliance on manual annotation while enhancing scalability, making it an attractive approach for various
surface defect identification tasks.