Few-shot anomaly detection using positive unlabeled learning with cycle consistency and co-occurrence features

Sion An, Jaehong Kim, Soopil Kim, Philip Chikontwe, Jiwook Jung, Hyejeong Jeon, Sang Hyun Park

Abstract: Highlights•We propose a few-shot anomaly detection method using positive and unlabeled learning.•Two key criteria are proposed for reliable positive and unlabeled learning.•We achieve state-of-the-art performance on three anomaly detection datasets.•Our method is robust in the composition of the unlabeled data and hyper-parameters.