Image Entropy Equalization for Autoencoder-Based One-Class Classification

Abstract

Autoencoder (AE) is a common technique for one-class classification (OCC). Reconstruction error (RE) is used to classify one seen class or other unseen classes. However, AE-based OCC (OCAE) does not provide a high AUC score. This study considers the hypothesis that RE is related to image entropy, and the OCAE is biased due to the image entropy differences. Based on such a hypothesis, this paper proposes image entropy equalization as the preprocessing technique. In which, image pixels are replaced by a defined set of pixels. Entropy equalized images are experimented with OCAE using MNIST, Fashion MNIST, and CIFAR10 datasets. Image Entropy Equalization improves AUC scores with several seen classes, where the improved classes have relatively high entropy on original images.