Design of novel generative adversarial network (GAN) architecture and its application for video anomaly detection

Abstract

Generative Adversarial Network (GAN) is an advanced deep learning method consisting of the Generator and Discriminator network. It offers the advantage over other deep-learning methods such as Deep Convolutional Neural Networks (DCNNs) and Auto-Encoder (AEs) for generating highly realistic and diverse data by employing adversarial training between the Generator and Discriminator networks. It can effectively model the single class by generating realistic samples that resemble the training data. This approach enhances the detection of anomalies by identifying deviations from the generated normal patterns. Thus, GAN methods are incorporated for video Anomaly Detection (AD) problems where the role of the Generator network is to learn and generate the normal video frames, while the role of the Discriminator network is to distinguish between the generated video frame and the real video frame. The generation of the video frame can be categorized as a reconstruction based method and a future frame prediction method. In the reconstruction-based method, the Generator network reconstructs the input video frame while in the future frame prediction method, the Generator network predicts the future video frame from the input video frame or the sequence of input video frames.