Unsupervised-Anomaly-Detection-with-SSIM-AE

Story on Medium

Anomaly Detection in Computer Vision with SSIM-AE

Related Works

[1] Bergmann, Paul, et al. "Improving unsupervised defect segmentation by applying structural similarity to autoencoders." arXiv preprint arXiv:1807.02011 (2018)

[2] Bergmann, Paul, et al. "MVTec AD--A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019

Data

Used datasets are available on the websites:

• grid and carpet: https://www.mvtec.com/company/research/datasets/mvtec-ad/,
• woven fabrics (texture_1 and texture_2): https://www.mvtec.com/company/research/publications/

Put downloaded datasets in directory data/

Training

python AE_training.py [-h] [--dataset_name DATASET_NAME] [--latent_dim LATENT_DIM] [--batch_size BATCH_SIZE] [--training_loss TRAINING_LOSS] [--load_model LOAD_MODEL] [--random_crop RANDOM_CROP]

Parameters:

• dataset_name (name of dataset used for training) e.g. "grid", "carpet", "texture_1", "texture_2",
• latent_dim (dimension of bottleneck in autoencoder architecture) e.g. 100,
• batch_size (batch size used for autoencoder training) e.g. 8,
• training_loss (loss used for autoencoder training): "ssim" or "mse",
• load_model (load weights of trained model): 1 or 0,
• random_crop (random crop 10k ROIs of size 128): 1 or 0.

Evaluation

python AE_evaluation.py