• Andreas Bach Daasbjerg
Anomaly detection can be seen as a one-class classification problem, due to the rare occurrence of anomalous frames.
Introducing pseudo-anomalies to the training set can potentially improve the performance of the anomaly detection model.
This project introduces a pipeline for improving the unsupervised anomaly detection task, by teaching a generator to generate pseudo-anomalies.
The pseudo-anomalies are generated by increasing a loss component in the overall loss function of the generator. Two loss components were tested, kullback-leibler divergence and flow loss.
The pseudo-anomalies are used to train a classifier to classify normal and abnormal frames.
Upon evaluation of the model, an AUC score is calculated using a combination of the classification score and the psnr score.
The pipeline achieved an AUC-score of 72.42\% evaluated on the CUHK Avenue dataset.
While not achieving state-of-the-art results, the pipeline shows potential for improving the performance of anomaly detection tasks.
Future work could include adding a second generator branch to generate normal frames.
Another approach is to evaluate the performance of the pipeline on a different dataset, such as the ShanghaiTech dataset.
Publication date1 Jun 2023
Number of pages91
ID: 532632699