This is the code used to produce the main results in the paper:
Sarafijanovic-Djukic N, Davis J. Fast Distance-Based Anomaly Detection in Images Using an Inception-Like Autoencoder. InInternational Conference on Discovery Science 2019 Oct 28 (pp. 493-508). Springer, Cham.
@inproceedings{sarafijanovic2019fast,
title={Fast Distance-Based Anomaly Detection in Images Using an Inception-Like Autoencoder},
author={Sarafijanovic-Djukic, Natasa and Davis, Jesse},
booktitle={International Conference on Discovery Science},
pages={493--508},
year={2019},
organization={Springer}
}\
Usage example:
python main.py --dataset cifar100 --normal_class 11 --features_extractor cae --cae_type inception --anomaly_detection qnnd --output_dir results2 --random_seed 15 --param_k 1 --param_m 2 --param_c 3
Parameters for main.py:
-d or --dataset: type of datasets, choices: ['mnist', 'fmnist', 'cifar10', 'cifar100'], required parameter
-nc or --normal_class: class for normal images, choices: if dataset is 'cifar100' it is in range 0..19, otherwise it is in range 0..9", required parameter
-fe or --features_extractor - tells if raw image or low-dimensional representation obtained by convolutional auto-encoder (CAE) is used, choices: ['raw','cae'], default='cae'
-cae or --cae_type - the type of CAE, choices=['baseline','inception']; default='inception'
-ad or --anomaly_detection- anomaly detection method: ocsvm - One Class Support Vector Machines, nnd - exact distance-based, qnnd - approximated distance-based with product quantization", choices=['ocsvm','nnd', 'qnnd'], default='qnnd'
-odir or --output_dir- directory to store the output results, required parameter
-rs or --random_seed - random seed, not required, default=0
-k or --param_k - nnd and qnnd parameter k, default=1
-m or --param_m - qnnd parameter m, choices = [1,2,4,8,16,32,64,128], default=1
-c or --param_c - qnnd parameter c, choices = [1,2,3,4,5,6,7,8], default=1