Positive and unlabeled learning on generating strategy for weakly anomaly detection

Shizhuo Deng, Bowen Han, Xiaohong Li, Siqi Lan, Dongyue Chen, Tong Jia, Hao Wang

Abstract: Anomalies are rare, contextual, and hard to annotate in anomaly detection scenarios. Usually, anomalies are coarse-grained labeled and there exists at least one abnormal patch in image and video segmentation. In such a setting, anomaly detection encounters anomaly diversity, quantity, and weakly-label problems. In this work, we propose a weakly anomaly detection framework by using Positive-unlabeled Learning (PUL) on generated weakly-labeled samples. This framework consists of sample generation module and anomaly detection module. In sample generation, Wasserstein Generative Adversarial Networks (WGAN) generates unlabeled patch-level samples including abnormal and normal classes. The unlabeled generated samples increase the quantity and diversity of anomalies for the PUL task. In anomaly detection model, a coarse-grained two-stage iterative PUL algorithm is proposed on positive labeled normal samples and unlabeled generated samples, which improves accuracy and stability. In the experiment, we evaluate our proposal in ablation study, which verifies the effectiveness of combining WGAN and PUL. Compared with typical anomaly detection models, our framework performs better with enhancements of 1.09% and 0.8% on the UCSD-Ped2 and Avenue datasets.