Learning Graph-Fourier Spectra of Textured Surface Images for Defect Localization

Abstract

This project addresses the challenge of detecting defects in images with highly textured backgrounds, common in industrial manufacturing. By leveraging graph Fourier analysis (Sandryhaila & Moura, 2014), the approach identifies defective images and crucial graph Fourier coefficients that highlight defects. Our 1D-CNN architecture achieves 99.4% classification accuracy with SHAP values (Lundberg & Lee, 2017) providing interpretable insights into spectral coefficient importance.

Installation

1. Clone the repository:

   git clone https://github.com/ganatma/GFT-based-Anomaly-Localization.git
   cd GFT-based-Anomaly-Localization

2. Install requirements:

   pip install -r requirements.txt

Dataset

The MVTec AD dataset (Bergmann et al., 2019) is used for evaluation. To download:

1. Visit MVTec AD dataset portal
2. Complete the academic license agreement
3. Use provided download scripts

Preprocess with dataset_creation.ipynb to generate patched datasets.

Methodology

Feature Extraction

*Graph Fourier Transform pipeline (Sandryhaila & Moura, 2014) for texture representation*

Classification Pipeline

*End-to-end defect detection workflow*

Results

Classification Performance

Class	Mean Pixel-wise AUROC (%)
Wood	90.79
Tile	93.50
Leather	98.40
Carpet	93.20
Grid	93.20

Classification Performance

*Convergence plot showing 99.94% training accuracy and 99.91% validation accuracy over 50 epochs*

Model Interpretation

*SHAP values identifying critical low-frequency components (Lundberg & Lee, 2017)*

Benchmark Comparison

*State-of-the-art comparison on MVTec AD (Bergmann et al., 2019)*

Citation

If you use this work in your research, please cite:

@article{nakkina2024learning,
  title={Learning graph-Fourier spectra of textured surface images for defect localization},
  author={Nakkina, Tapan Ganatma and Karthikeyan, Adithyaa and Eksin, Ceyhun and Bukkapatnam, Satish TS},
  journal={Manufacturing Letters},
  volume={41},
  pages={1568--1578},
  year={2024},
  publisher={Elsevier}
}

References

• Bergmann, P., et al. (2019). MVTec AD -- A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection. CVPR.
• Kiranyaz, S., et al. (2021). 1D Convolutional Neural Networks for Signal Processing Applications. IEEE TII.
• Lundberg, S. M., & Lee, S.-I. (2017). A Unified Approach to Interpreting Model Predictions. NeurIPS.
• Nakkina, T. G., et al. (2024). Learning graph-Fourier spectra of textured surface images for defect localization. Manufacturing Letters, 41, 1568-1578.
• Sandryhaila, A., & Moura, J. M. (2014). Discrete Signal Processing on Graphs. IEEE TSP.

Contributing

This project adheres to the Contributor Covenant. Please see CONTRIBUTING.md for details.

License

MIT License - See LICENSE for full text.