This repository presents a patch-wise statistical feature encoding (PSFE) method for pixel-level visual anomaly detection, motivated by the failure of global feature aggregation to capture localized structural defects.
This repository is intended as a research prototype and companion implementation for patch-level statistical anomaly detection.
The project includes:
- A baseline global feature model highlighting its limitations.
- A patch-wise statistical encoding approach for anomaly localization.
- Quantitative and qualitative evaluation on the MVTec AD dataset.
- Motivation
- Dataset
- Repository Structure
- Baseline: Global Feature Aggregation
- Proposed Method: PSFE
- Pixel-level Localization Performance
- Quantitative Evaluation
- Ablation Study
- Key Observations
- Future Work
- Acknowledgements
Most deep feature-based anomaly detection methods rely on global average pooled representations, which suppress localized defect cues.
We show that such global representations perform poorly on structural anomalies and propose a patch-level statistical alternative that requires no training and no memory bank.
- MVTec Anomaly Detection (AD) Dataset
- Categories used:
- Carpet
- Grid
- Bottle
- Capsule
- The training set contains only normal samples, following the standard MVTec AD protocol.
- Evaluation performed at pixel-level using ground-truth segmentation masks.
Dataset: MVTec Anomaly Detection (AD)
The repository is organized as follows:
.
├── checkpoints/
│ └── baseline_shallow_autoencoder.pth
│
├── notebooks/
│ ├── 01_baseline_global_ae.ipynb
│ └── 02_psfe.ipynb
│
├── evaluation_visuals/
│ ├── anomaly_score_distribution(baseline).png
│ ├── baseline_roc_curve.png
│ ├── Pixel-level-roc-curve.png
│ └── psfe-heatmaps.png
│
├── data/
| ├── carpet
| ├── grid
| ├── bottle
| ├── capsule
│ └── README.md
│
├── requirements.txt
├── research_paper_reference.pdf
└── README.md
- Extract deep features using ResNet-50
- Intermediate layers: Layer2 and Layer3
- Apply Global Average Pooling (GAP)
- Train a shallow Autoencoder on global features
- Use reconstruction error as anomaly score
- Perform image-level anomaly classification
Global feature aggregation suppresses localized defect patterns, leading to poor separability of structural anomalies.
Baseline ROC-AUC ≈ 0.51
Anomaly Score Distribution (Baseline)
.png)
Baseline ROC Curve
Instead of aggregating features globally, anomalies are detected by statistical deviations at the patch level.
- Extract ResNet-50 Layer2 & Layer3 feature maps
- Spatially align and concatenate feature maps
- Divide features into patch-wise representations
- Compute non-parametric statistics:
- Patch-level variance
- Patch-level entropy
- Generate pixel-level anomaly maps
- Aggregate maps to obtain anomaly scores
No Autoencoder
No PaDiM
No PatchCore
Training-free inference
PSFE Anomaly Heatmaps
Pixel-level ROC Curve
Pixel-level AUROC scores are reported across selected MVTec categories.
| Category | Mean Pixel AUROC | Std |
|---|---|---|
| Carpet | 0.989 | 0.005 |
| Grid | 0.934 | 0.077 |
| Bottle | 0.963 | 0.041 |
| Capsule | 0.944 | 0.053 |
| Method | Mean Pixel AUROC | Std |
|---|---|---|
| PSFE (variance only) | 0.989488 | 0.005129 |
| PSFE (variance + entropy) | 0.989431 | 0.005169 |
Including entropy does not significantly affect AUROC but improves stability of patch-wise responses, especially in structurally repetitive textures.
- Global representations fail to capture localized anomalies
- Patch-wise statistics provide strong localization cues
- Training-free approaches can be competitive with learning-based anomaly detection methods
- Statistical encoding offers interpretability and simplicity
- Extension to additional MVTec categories
- Multi-scale patch aggregation
- Evaluation on real-world industrial datasets
- Comparison with memory-bank-based methods
- Pretrained ResNet-50 from torchvision
- MVTec AD Dataset