Multi-grained teacher–student joint representation learning for surface defect classification

Abstract

Surface defect classification (SDC) plays a critical role in ensuring product quality within industrial systems. Surface defects are characterized by complex noise backgrounds, diverse defect types, and multi-scale defect shapes. Existing methods often struggle to effectively learn multi-grained defect information in such complex environments. This study introduces a Multi-Grained Teacher-Student Joint Representation Learning (MGJR) framework, which integrates both coarse-grained and fine-grained representation learning in a unified architecture. A ViT-based teacher network first learns holistic global features from defect-rich backgrounds. These features guide a student network enhanced with an Integrated Efficient Multi-Attention (IEMA) module and a Global-Local Attention (GL-Attention) mechanism, enabling the extraction and fusion of multi-scale features to preserve context while emphasizing local anomalies. Additionally, the anchor-guided training strategy (AGTS) serves as a consistency constraint, enhancing robustness by aligning the teacher's stable coarse-grained signal with the student model's fine-grained response under noisy inputs. The entire framework is optimized end-to-end using a unified loss that combines coarse-level guidance with task-specific supervision. Extensive experiments demonstrate that MGJR achieves 99.98% accuracy on the NEU-CLS dataset and consistently outperforms previous methods across multiple industrial benchmarks. The model remains lightweight, with 21.14 million parameters and 2.86 billion FLOPs. MGJR shows good performance in noisy conditions and other classification tasks. To demonstrate its practical effectiveness, this study built a wood surface defect dataset with 7 defect types and 2,654 images from real industrial settings. MGJR achieved top performance on this dataset, verifying its applicability in real-world.