Page 53 - Read Online

Hu et al. J. Mater. Inf. 2025, 5, 44 https://dx.doi.org/10.20517/jmi.2025.21 Page 3 of 17

learning has proven effective in addressing metal defect detection challenges in datasets with limited
[33]
labeling . Self-supervised pre-training models are typically built on contrastive learning frameworks [34,35] ,
which train models to differentiate data samples by comparing similarities and differences between data
[36]
points. Representative methods include momentum contrast (MOCO) , simple framework for contrastive
learning of visual representations (SimCLR) , and simple Siamese network (SimSiam) . MOCO
[37]
[38]
optimizes contrastive learning by dynamically managing a large number of negative samples using a
dictionary queue, though its complex architecture demands significant computational resources. SimCLR
enhances performance through large batch sizes and advanced data augmentation techniques to generate
negative sample pairs but similarly requires substantial computational resources and carefully designed
augmentation strategies. In contrast, SimSiam stands out for its simplicity and efficiency. It eliminates the
need for negative samples or momentum encoders and prevents feature collapse via stop-gradient
operations and symmetric predictor designs, resulting in a streamlined architecture that is easier to
implement and scale. SimSiam has achieved state-of-the-art performance across multiple benchmark
datasets.

In this study, we introduce a streamlined self-supervised defect detection framework and devise a weight
transfer scheme to pre-train the comparative learning SimSiam model on a comprehensive dataset of
unlabeled images to learn their inherent features. Subsequently, we employ the learned weights of the model
to serve as a feature extractor within Faster R-CNN for application on a constrained dataset of labeled steel
surface defect images to evaluate detection performance. This study examines the factors influencing
detection accuracy and confirms the efficacy of self-supervised learning approaches in steel surface defect
detection. The principal contributions of this paper are as follows:

(1) This paper introduces a novel self-supervised learning approach specifically tailored for steel surface
defect detection. It utilizes unlabeled data to train a model capable of generalizing effectively to various
types of defects with a high degree of accuracy. The approach markedly diminishes dependence on manual
labeling, which is labor-intensive and costly, and offers a scalable solution for microstructural image
classification and localization in surface defect identification.

(2) We propose a streamlined self-supervised learning framework for steel surface defect detection that
reduces model complexity, enhances interpretability and reliability, negates the need for extensive labeled
data, and abbreviates detection time.

(3) Extensive experiments demonstrate that our approach achieves superior results compared to a baseline
model using random weights and ImageNet pre-trained ResNet18 weights on the publicly available
downstream defect dataset NEU-DET. Our study underscores the viability of self-supervised methods in
this domain and lays the groundwork for more advanced defect detection techniques.

The subsequent sections of this paper are organized as follows: The “Methodology” section details the
proposed methodology, including a comprehensive description of dataset creation and the self-supervised
learning framework. The “Results” section delineates the experimental results and analysis. The
“Discussion” section discusses the limitations of the study and proposes future research directions. Finally,
the “Conclusions” section encapsulates the findings of the paper.