ADGAN: A Scalable GAN-based Architecture for Image Anomaly Detection

With deep learning booming, related technologies have been applied in various fields. However, it remains an open question in terms of how to perform well on anomaly detection of images with diverse content and complexity. To address such problem, we propose ADGAN (Anomaly Detection Generative Adversarial Network), a scalable encoder-decoder-encoder architecture for image anomaly detection. Through extracting and utilizing multi-scale features of normal samples, we obtain fine-grained reconstructed images of normal class. Combined with adversarial training, the proposed model learns the distribution of normality thus large reconstruction errors occur when it processes anomalous samples during inference. We verify the effectiveness of ADGAN on two benchmark datasets: CIFAR-10 and CIFAR-100. The experimental results demonstrate that our method outperforms current anomaly detection work. We improve the top performing baseline AUCs by 9% and 6% on the CIFAR-10 dataset and the CIFAR-100 dataset respectively.