Pseudo-anomaly generation for improving the unsupervised anomaly detection task.: Implementation of a generative neural network
Author
Daasbjerg, Andreas Bach
Term
4. term
Publication year
2023
Submitted on
2023-06-01
Pages
91
Abstract
Det er svært at opdage afvigelser (anomalier), fordi de næsten aldrig forekommer. Derfor behandles opgaven som en én-klasses klassifikation, hvor man mest har eksempler på normale billeder (frames). Dette projekt foreslår en pipeline, der forbedrer usuperviseret anomali-detektion ved at lære en generator at skabe pseudo-anomalier, dvs. kunstigt fremstillede afvigelser. Generatoren tilskyndes til at lave mere unormale billeder ved bevidst at øge en valgt del af dens tabsfunktion. To sådanne tabskomponenter blev afprøvet: Kullback-Leibler-divergens (et mål for forskel mellem sandsynlighedsfordelinger) og flow loss. De genererede pseudo-anomalier bruges derefter til at træne en klassifikator, der skelner mellem normale og unormale frames. Ved evaluering beregnes en AUC-score ved at kombinere klassifikatorens score med en PSNR-score (Peak Signal-to-Noise Ratio, et mål for hvor forskellige to billeder er). På CUHK Avenue-datasættet opnår pipelinen en AUC på 72,42%. Resultatet er ikke state of the art, men viser potentiale for at forbedre anomali-detektion. Fremtidigt arbejde kan omfatte en ekstra generatorgren til at skabe normale billeder eller evaluering på andre datasæt som ShanghaiTech.
Detecting anomalies is challenging because they are rare, so the task is often treated as one-class classification with many examples of normal frames but few or none of abnormal ones. This project proposes a pipeline to improve unsupervised anomaly detection by training a generator to create pseudo-anomalies, i.e., artificially produced outliers. The generator is encouraged to produce more abnormal outputs by deliberately increasing a chosen component of its loss function. Two components were tested: Kullback-Leibler divergence (a measure of difference between probability distributions) and flow loss. The generated pseudo-anomalies are then used to train a classifier to distinguish normal from abnormal frames. For evaluation, an AUC score is computed by combining the classifier's score with a PSNR score (Peak Signal-to-Noise Ratio, a measure of how different two images are). On the CUHK Avenue dataset, the pipeline achieves an AUC of 72.42%. While not state of the art, the results show potential for improving anomaly detection. Future work could add a second generator branch to produce normal frames or test the approach on other datasets such as ShanghaiTech.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Keywords
Documents