Robust multi-view anomaly detection method based on masking reconstruction and diversified reasoning

Abstract

The invention discloses a robust multi-view anomaly detection method based on masking reconstruction and diversified reasoning, and belongs to the technical field of anomaly detection. Aiming at the problems of data prior dependence, pollution data processing and cross-view correlation utilization insufficiency of the existing method, the method comprises the following steps: (1) introducing a learnable masking reconstruction mechanism, adaptively shielding part of features through a network, reconstructing residual features, strengthening the modeling capability of a feature dependence relationship, and realizing attribute anomaly detection; (2) designing diversified reasoning tasks among multiple views, and modeling cross-view dependence in combination with a masking diversity mechanism for detecting category anomaly; (3) introducing a loss-driven reweighting strategy for enhancement, and dynamically reducing the interference of abnormal samples on the model; experimental results show that the detection performance of the method is superior to that of an existing method on a plurality of real multi-view data sets, and the method shows higher stability and robustness under different pollution proportions. The method can be widely applied to anomaly detection tasks of various multi-view scenes.

Images