Industrial Surface Defect Detection via Diffusion Generation and Asymmetric Student-Teacher Network
Summary: arXiv:2604.19240v1 Announce Type: new
Abstract
Industrial surface defect detection often suffers from limited defect samples, severe long-tailed distributions, and difficulties in accurately localizing subtle defects under complex backgrounds. To address these challenges, this paper proposes an unsupervised defect detection method that integrates a Denoising Diffusion Probabilistic Model (DDPM) with an asymmetric teacher-student architecture.
Methodology Overview
The proposed approach tackles the issue of data scarcity and enhances defect detection accuracy through two main strategies:
- Data Level Enhancement: The DDPM is trained only on normal samples. By introducing constant-variance Gaussian perturbations and Perlin noise-based masks, the model generates high-fidelity and physically consistent defect samples along with pixel-level annotations. This effectively alleviates the data scarcity problem.
- Model Level Architecture: An asymmetric dual-stream network is constructed, where the teacher network provides stable representations of normal features. Conversely, the student network reconstructs normal patterns and amplifies discrepancies between normal and anomalous regions.
Optimization Strategy
A joint optimization strategy combining cosine similarity loss and pixel-wise segmentation supervision is employed to achieve precise localization of subtle defects. This approach ensures that the model is not only trained to identify anomalies but also to localize them accurately within the context of complex industrial surfaces.
Experimental Results
Experimental results on the MVTecAD dataset demonstrate the effectiveness of the proposed method. Key findings include:
- Achieving a 98.4% image-level Area Under the Receiver Operating Characteristic (AUROC).
- Attaining a 98.3% pixel-level AUROC.
- Significantly outperforming existing unsupervised and mainstream deep learning methods in defect detection accuracy.
Conclusion
The innovative approach presented in this study addresses significant challenges in industrial surface defect detection. By leveraging a combination of data generation techniques and a robust asymmetric teacher-student network, the method demonstrates the potential for accurate and efficient defect detection without requiring large amounts of real defect samples. This advancement paves the way for more reliable industrial inspection processes, thereby enhancing quality control and reducing production costs.
Keywords
Industrial defect detection, diffusion models, data generation, teacher-student architecture, pixel-level localization.