Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Quantized FCA: Efficient Zero-Shot Texture Anomaly Detection

Andrei-Timotei Ardelean, Patrick Rückbeil, Tim Weyrich

TL;DR

This work tackles zero-shot texture anomaly localization by exploiting internal texture statistics. It introduces QFCA, a quantized variant of FCA that represents patches as histograms over $N$ quantization levels and computes an anomaly map via a linear-time, GPU-accelerated transport inspired by the Wasserstein distance, achieving a 10x speedup with minimal accuracy loss. To further boost performance on complex textures, the authors add a PCA-based residual feature preprocessing, yielding QFCA+ that improves detection precision. Extensive experiments on MVTec AD textures, DTD-Synthetic, and WFT demonstrate real-time performance with competitive or state-of-the-art localization metrics. The work offers practical impact for industrial texture monitoring and provides open-source code for reproducibility.

Abstract

Zero-shot anomaly localization is a rising field in computer vision research, with important progress in recent years. This work focuses on the problem of detecting and localizing anomalies in textures, where anomalies can be defined as the regions that deviate from the overall statistics, violating the stationarity assumption. The main limitation of existing methods is their high running time, making them impractical for deployment in real-world scenarios, such as assembly line monitoring. We propose a real-time method, named QFCA, which implements a quantized version of the feature correspondence analysis (FCA) algorithm. By carefully adapting the patch statistics comparison to work on histograms of quantized values, we obtain a 10x speedup with little to no loss in accuracy. Moreover, we introduce a feature preprocessing step based on principal component analysis, which enhances the contrast between normal and anomalous features, improving the detection precision on complex textures. Our method is thoroughly evaluated against prior art, comparing favorably with existing methods. Project page: https://reality.tf.fau.de/pub/ardelean2025quantized.html

Quantized FCA: Efficient Zero-Shot Texture Anomaly Detection

TL;DR

This work tackles zero-shot texture anomaly localization by exploiting internal texture statistics. It introduces QFCA, a quantized variant of FCA that represents patches as histograms over quantization levels and computes an anomaly map via a linear-time, GPU-accelerated transport inspired by the Wasserstein distance, achieving a 10x speedup with minimal accuracy loss. To further boost performance on complex textures, the authors add a PCA-based residual feature preprocessing, yielding QFCA+ that improves detection precision. Extensive experiments on MVTec AD textures, DTD-Synthetic, and WFT demonstrate real-time performance with competitive or state-of-the-art localization metrics. The work offers practical impact for industrial texture monitoring and provides open-source code for reproducibility.

Abstract

Zero-shot anomaly localization is a rising field in computer vision research, with important progress in recent years. This work focuses on the problem of detecting and localizing anomalies in textures, where anomalies can be defined as the regions that deviate from the overall statistics, violating the stationarity assumption. The main limitation of existing methods is their high running time, making them impractical for deployment in real-world scenarios, such as assembly line monitoring. We propose a real-time method, named QFCA, which implements a quantized version of the feature correspondence analysis (FCA) algorithm. By carefully adapting the patch statistics comparison to work on histograms of quantized values, we obtain a 10x speedup with little to no loss in accuracy. Moreover, we introduce a feature preprocessing step based on principal component analysis, which enhances the contrast between normal and anomalous features, improving the detection precision on complex textures. Our method is thoroughly evaluated against prior art, comparing favorably with existing methods. Project page: https://reality.tf.fau.de/pub/ardelean2025quantized.html
Paper Structure (24 sections, 8 equations, 10 figures, 7 tables, 1 algorithm)

This paper contains 24 sections, 8 equations, 10 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Algorithm Design
  4. Patch Statistics Comparison
  5. Reference Selection
  6. Feature Extractor
  7. Experiments
  8. Datasets
  9. Metrics
  10. Comparison with prior art
  11. Image and patch size
  12. In-depth ablation study
  13. Number of Bins
  14. Feature preprocessing
  15. Computing the Reference
  16. ...and 9 more sections

Figures (10)

  • Figure 1: Anomaly localization fidelity (PRO) vs Time, obtained using different algorithms at various resolutions. Our approach enables the best tradeoff between accuracy and runtime.
  • Figure 2: Our QFCA pipeline: each patch is represented as a histogram and compared to the reference using our Algorithm \ref{['algorithm']}. The final score for each pixel is given by averaging the mismatch scores from the histograms it is part of.
  • Figure 3: Runtime for average pooling with different kernel sizes, operating on a 128×128 tensor with 8192 channels (mirroring our 16 bins per feature setting). All methods use GPU acceleration.
  • Figure 4: Visualization of anomaly localization results on challenging examples by different methods.
  • Figure 5: Diagram showing the pro score and execution time with patch sizes 3, 5, 7, 9, or 11. The optimum patch-size depends on the image size.
  • ...and 5 more figures