A Clustering-Based Deep Autoencoder For One-Class Image Classification

Anomaly detection in images is a topic of great interest in Computer Vision. It can be defined as an One-Class problem, where the goal is to detect deviations from known patterns, which are defined as normal. Recently, Deep Learning methods became popular due to their performance on classification tasks. This works presents an image anomaly detection classifier based on a previously known method, the Deep Embedded Clustering, which is based on a Deep Autoencoder. We show the effectiveness of the method through three different experiments. Results suggest that the method improves classification performance when compared to a Stacked Denoising Autoencoder in the image anomaly detection context.