Autoencoders vs. others for anomaly detection
| Název česky | Autoencodery vs. jiné metody pro detekci anomálií |
|---|---|
| Autoři | |
| Rok publikování | 2018 |
| Druh | Článek ve sborníku |
| Konference | DATA A ZNALOSTI & WIKT 2018, sborník konference |
| Fakulta / Pracoviště MU | |
| Citace | |
| Klíčová slova | Autoencoders · Local Outlier Factor (LOF) · z-score · anomalies · PascalVOC image dataset · feature extraction |
| Popis | The paper deals with a task of finding anomalies in the set of pictures using autoencoders and comparison of results with other methods searching for outliers, namely LOF and z-score. Outliers found by these methods are compared to outliers found by project team members. Process consists of preprocessing of pictures using pretrained deep neural nets (one at a time), reducing dimension using PCA, normalization of features and applying methods on pictures, either on a whole set or subsets divided by classes (dividing the pictures to groups by objects of interest that can be found in them). Output of methods with different attribute settings was compared to outliers found by team members using confusion matrix and F1-score. The results were not very positive, no significant relationships were found between anomalies found by team members and by anomalies found by individual methods. Possible reasons for this are discussed. |
| Související projekty: |