Abstraction-Based Outlier Detection for Image Data
Source of Publication
Advances in Intelligent Systems and Computing
© 2021, Springer Nature Switzerland AG. Data plays an important role in all stages of training, and usage of machine learning algorithms. Outliers are the samples in data that are generated by a “different mechanism” and belong to unexpected patterns that do not conform to normal behaviour. Outlier detection techniques try to deal with such undesirable events. There have been exceptional success of deep learning over classical methods in computer vision. In recent years a number of works employed the representation learning ability of deep autoencoders or Generative Adversarial Networks for outlier detection. Basically, methods are based on plugging representation techniques to outlier detection methods or directly reported employing reconstruction error as an outlier score. The error distributions of inliers and outliers may be still significantly overlapped. This could be associated with variation of samples inside the class, or cases with high outliers ratios, etc. In these cases, simply thresholding reconstruction errors may lead to misclassification. Although the produced representation is perhaps effective in representing the common features of the normal data, it is not necessarily effective in distinguishing outliers from inliers. We present a method that is based on constructing new features using convolutional variational autoencoder (VAE) and generate abstraction based on these features. To identify anomaly detection we tested two scenarios: utilizing VAE itself as well as using abstractions to train an additional architecture. Results are presented in the form of AUC-ROC using four benchmark datasets.
Yakovlev, Kirill; Bekkouch, Imad Eddine Ibrahim; Khan, Adil Mehmood; and Khattak, Asad Masood, "Abstraction-Based Outlier Detection for Image Data" (2021). Scopus Indexed Articles. 10.