Face Validation Based Anomaly Detection Using Variational Autoencoder

In unconstrained facial images, large visual variations concerning pose, scale, the presence of occlusions, expressions and lighting usually cause difficulties in discriminating faces from the background accurately. As a result, some non-face regions are recognized as faces (false positive) and that influences the effectiveness of face detection algorithms which is characterized by low false positive (FP) rate, high detection rate and high speed of processing. In order to reduce these non-face regions, they are considered as anomalies and then try to detect them. In this paper, we propose an anomaly detection method using reconstruction error from variational autoencoder (VAE), which is a generative machine learning model. We train VAE to learn reconstructing faces that are close to its original input faces using FDDB dataset, then the difference between the original input face and the reconstructed output is measured to obtain the reconstruction error which can be used as an anomaly score. Consequently, the regions resulting from faces detection algorithm with high reconstruction error are defined as anomalies or false positives.