Page 54 - Read Online

Page 4 of 17 Hu et al. J. Mater. Inf. 2025, 5, 44 https://dx.doi.org/10.20517/jmi.2025.21

MATERIALS AND METHODS
Dataset acquisition and data pre-processing
[39]
In this study, a strip steel surface defect image dataset provided by Northeastern University (NEU) was
used. The dataset contains six major categories of surface defects associated with hot-rolled strips: slag (RS),
indentation (Pa), crack (Cr), speckle (PS), inclusions (In), and scratches (Sc). Examples of six defect types
are shown in Figure 1A. 300 samples were taken from each category, resulting in a comprehensive dataset of
1,800 grayscale images. Each image in the NEU dataset has a resolution of 200 × 200 pixels. In the defect
detection task, the NEU-DET dataset also provides bounding box annotations to delineate the defect
categories and their spatial locations in each image. The bounding boxes in the NEU-DET dataset were
exclusively utilized during fine-tuning as annotation labels. The dataset is available on the web (http://
faculty.neu.edu.cn/songkechen/zh_CN/zdylm/263270/list/index.htm).

To further expand the unlabeled dataset for self-supervised learning, we acquired the dataset from the
Severstal Steel Defect Detection competition hosted on the Kaggle platform. The dataset is available on the
web (https://www.kaggle.com/competitions/severstal-steel-defect-detection/data). Original training images
(1,600 × 256 pixels) were cropped along the length direction into four sub-images of equal dimensions. This
process generated augmented samples while preserving defect features. The merged dataset (hereafter
referred to as the SSDD dataset) consists of 20,272 images of surface defects on strips, each 400 × 256 pixels
in size. Examples of these defects are shown in Figure 1B.

Following the specified protocol, we integrate the NEU dataset with the SSDD dataset to form a new pre-
training dataset named SSDD + NEU.

Self-supervised learning framework
Self-supervised learning has emerged as a promising approach for advancing machine learning. Unlike
supervised learning, which requires labeled data, self-supervised learning generates supervisory signals
directly from unannotated visual data, enabling the acquisition of generalizable and transferable
[40]
representations without human intervention . In computer vision, both transfer learning and self-
supervised learning are employed for model pre-training . Critically, self-supervised learning facilitates
[41]
pre-training on large-scale datasets of unlabeled images, substantially expanding data resources. Figure 2
illustrates our workflow for self-supervised learning-enhanced material defect detection.

The algorithm comprises a pre-training stage followed by a downstream defect detection task. During self-
supervised pre-training, the model learns from unlabeled images of steel surface defects. This phase focuses
not on direct defect detection but on learning generic image representations. The resulting weights
encapsulate transferable features, providing a foundation for downstream tasks. In transfer learning, these
weights serve as feature extractors for steel defect detection. Typically, only the final layers require fine-
tuning when adapting to this task: deeper layers retain general features, while output-proximal layers
necessitate task-specific adaptation. By fine-tuning on limited labeled defect data, the model acquires defect-
specific characteristics. During detection, the model processes new steel surface images to localize and
classify defects. Leveraging rich pre-trained features and task-adapted fine-tuning enables effective defect
detection with minimal labeled data. The self-supervised pre-training employs an online augmentation
pipeline adapted from established contrastive learning methodologies. For each input image, two
stochastically augmented views are generated via random resized cropping, color jittering, random grayscale
conversion, gaussian blurring, and horizontal flipping. This multi-transformation strategy ensures robust
feature learning while preserving defect discriminability. Critically, independent augmentation sampling per
view maintains diversity and prevents trivial solutions during representation learning.