MLF SC Save
Anomaly detection method that incorporates multi-scale features to sparse coding
MLF-SC: Incorporating Multi-Layer Features to Sparse Coding for Unsupervised Anomaly Detection
MLF-SC (Multi-Layer Feature Sparse Coding) is an anomaly detection method that incorporates multi-scale features to sparse coding. This is a PyTorch implementation for MVTec datasets (Carpet, Grid, Leather, Tile, Wood, ...).
Output visualization of MVTec dataset (Grid, Hazelnut, and Bottle)
Quick Start
git clone [email protected]:LeapMind/MLF-SC.git
pip3 install -r requirements.txt
python3 main.py train cfg/sample_config.yml
python3 main.py test cfg/sample_config.yml
Download Dataset
wget ftp://guest:[email protected]/mvtec_anomaly_detection/mvtec_anomaly_detection.tar.xz
tar -xvf mvtec_anomaly_detection.tar.xz
Usage
You can train and test for each texture dataset.
Set dataset path to config.yml
Set root in config.yml to texture dataset path (like path/to/mvtec_anomaly_detection/carpet/).
Train
$ python3 main.py train cfg/config.yml
Test and Visualize Output
$ python3 main.py test cfg/config.yml
Contribution
Anomaly detection performance for the texture categories of the MVTec AD dataset. For each cell in the R1 / R2 columns, the ratio of correctly classified samples of normal R1 and that of anomalous images R2 are shown with R1 / R2 notation. The maximum averages (R1 + R2) / 2 are marked with boldface. The performance for the non-sparse-coding-based methods are cited from Table 2 of (Bergmann et al., 2019). The AUROC columns show only sparse coding and MLF-SC.
| R1 / R2 | AUROC | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Category | AE (SSIM) | AE (L2) | AnoGAN | CNN Feature Dictionary |
Texture Inspection |
Sparse Coding |
MLF-SC (Proposed) |
Sparse Coding |
MLF-SC (Proposed) |
| Carpet | 0.43 / 0.90 | 0.57 / 0.42 | 0.82 / 0.16 | 0.89 / 0.36 | 0.57 / 0.61 | 0.43 / 0.79 | 1.00 / 0.98 | 0.58 | 0.99 |
| Grid | 0.38 / 1.00 | 0.57 / 0.98 | 0.90 / 0.12 | 0.57 / 0.33 | 1.00 / 0.05 | 0.76 / 0.72 | 1.00 / 0.88 | 0.89 | 0.97 |
| Leather | 0.00 / 0.92 | 0.06 / 0.82 | 0.91 / 0.12 | 0.63 / 0.71 | 0.00 / 0.99 | 0.84 / 0.96 | 0.97 / 0.97 | 0.95 | 0.99 |
| Tile | 1.00 / 0.04 | 1.00 / 0.54 | 0.97 / 0.05 | 0.97 / 0.44 | 1.00 / 0.43 | 0.94 / 0.60 | 0.94 / 0.76 | 0.86 | 0.92 |
| Wood | 0.84 / 0.82 | 1.00 / 0.47 | 0.89 / 0.47 | 0.79 / 0.88 | 0.42 / 1.00 | 0.84 / 0.60 | 0.95 / 0.98 | 0.97 | 0.99 |
| Average | 0.53 / 0.74 | 0.64 / 0.65 | 0.90 / 0.18 | 0.77 / 0.54 | 0.60 / 0.62 | 0.７６ / 0.81 | 0.97 / 0.91 | 0.85 | 0.97 |
License
Non-commercial, research purposes only
License of dependent libraries
See LICENSE directory.
Open Source Agenda Badge
