Research on Scratch Defect Segmentation of Steel Surfaces Using MOSD-UNet Within a Semi-Supervised Self-Training Framework

To address the challenges of high annotation costs, class imbalance, and poor segmentation of small targets in steel scratch detection, we propose a semi-supervised self-training framework featuring Metal Object Scratch Detection U-Net (MOSD-UNet), a lightweight U-Net-based model. The framework initializes the model using limited labeled data and iteratively refines unlabeled data through high-confidence pseudo-labels, thereby reducing dependence on manual annotation. We designed a shallow feature-guided convolution module, termed Polarized Small Target Convolution (PST-Conv), which integrates the Channel Attention Module (CAM) and Gaussian Spatial Attention Module (GSAM) with a learnable Gaussian kernel to preserve fine-grained spatial features and mitigate information loss caused by pooling. Concurrently, we developed an efficient upsampling module called the Greatest Common Divisor Grouped Transposed Convolution (GCD-ConvT) module, which leverages grouped transposed convolution and channel compression to restore feature resolution while reducing parameters and enhancing inference speed. Finally, we introduce a specially designed mixed loss function that combines the Gradient Harmonizing Mechanism for Classification (GHM-C) and D[Formula: see text] loss, dynamically balancing the gradient distribution between easy and difficult samples and improving the structural consistency of segmented regions. All models, experimental configurations, and related code used in this study have been uploaded to the GitHub repository: https://github.com/Hayden-Mo/MOSD-UNet.git.