Unsupervised anomaly detection in images using attentional normalizing flows

Image-based anomaly detection has been widely used in practice, but it is still a challenging task due to the irregularity of anomalies. Existing representation-based methods have achieved high accuracy metrics in image-based anomaly detection, but they are weak in capturing anomalous regions, resulting in small inter-class variance between the latent distributions of normal and anomalous data. Furthermore, most current neural networks have low parameter efficiency, which limits their inference performance. To address these issues, we propose a novel normalizing flow called AFlow. Our network represents multi-scale features with multi-scale convolutions instead of inputs of different resolutions, significantly reducing the network parameters and improving the fine-grained representation in features. Then, an attention module is introduced into the normalizing flow to further select the multi-scale features along feature channels, making the network focus on relevant regions with more object details to refine the objective of maximum likelihood estimation. Experiments on two public datasets show that the proposed method can achieve more accurate prediction results and higher parameter efficiency than previous detection models: the average AUROC of the proposed method reaches 99.75% and 97.2% on MVTec AD and BTAD datasets, respectively.