Improved-GANomaly

This is an improved version GANomaly with keras from graduate-level course Machine Learning class.

Introduction

Instead of calculating the differences between the features from generated images by the encoder trained on generated samples and the features from testing images by the encoder trained on regular samples, we re-defined a new anomaly score, which calculates the differences between features from generated and testing images by the encoder trained on regular samples.

Dependencies

• tensorflow
• sklearn
• time
• matplotlib
• pytorch
• random
• cv2
• os
• glob
• numpy

Dataset

• The anomaly detection dataset for training our improved GANomaly can be downloaded from here).

Results:

We evaluated our results over three baseline models: AnoGAN, EGBAD, and the original GANomaly on the mvtec testing datasets.

1. The results after comparing our improved GANomaly with AnoGAN and EGBAD:

	AnoGAN	EGBAD	Ours
Accuracy	70.8 %	74.7 %	83.8 %
F1-Score	67.6 %	85.5 %	89.0 %
Inference Time	827.3 s	101.5 s	11.9 s

2. In the second experiments, we evaluated our method's performacne by comparing with the original GANomaly on the Wood, Screw, and Pill testing images from the mvtec datasets. The results are as the follows:

	Original GANomaly		Ours
	Accuracy	F1-Score	Accuracy	F1-Score
Wood	75.6 %	85.5 %	93.7 %	96.0 %
Screw	73.5 %	83.2 %	94.4 %	96.3 %
Pill	84.6 %	90.5 %	90.2 %	90.8 %

3. We also evaluated data augmentation on the improved GANomaly's performance:

	Original GANomaly		Ours
	Accuracy	F1-Score	Accuracy	F1-Score
Wood	90.8 %	94.4 %	93.7 %	96.0 %
Screw	86.2 %	82.6 %	94.4 %	96.3 %
Pill	84.1 %	90.5 %	90.2 %	90.8 %

🔨Getting Started:

1. Clone this github repo.

git clone https://github.com/Joannechiao18/Improved-GANomaly.git

2. Please cd to the Improved-GANomaly-main, and run the file GANomaly_copy.ipynb in Google Colab.

🔗 Citation

@article{DBLP:journals/corr/abs-1805-06725,
  author    = {Samet Akcay and
               Amir Atapour Abarghouei and
               Toby P. Breckon},
  title     = {GANomaly: Semi-Supervised Anomaly Detection via Adversarial Training},
  journal   = {CoRR},
  volume    = {abs/1805.06725},
  year      = {2018},
  url       = {http://arxiv.org/abs/1805.06725},
  eprinttype = {arXiv},
  eprint    = {1805.06725},
  timestamp = {Mon, 13 Aug 2018 16:46:23 +0200},
  biburl    = {https://dblp.org/rec/journals/corr/abs-1805-06725.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}