Multi -directional decision fusion for black-box source-free anomaly detection

With growing data-privacy concerns in industrial anomaly detection, source-free domain adaptation aims to transfer the pre-trained source model to the target domain without source data. However, the source model is often unavailable due to its commercial value. This paper investigates a more challenging yet practical problem, namely black-box source-free domain adaptation, where only the outputs of the source model and unlabeled target data are available. Specifically, our method comprises two key stages: self-learning based pseudo-label and cluster separation based classifier-hypothesis. In the first stage, we fuse instance-directional and class-directional decisions to generate pseudo-labels for target samples, transferring detection knowledge from the source model to the target model while mitigating the adverse effects of severe category-imbalance. Additionally, the spatial regularization is introduced to enhance the learning of discriminative target-features. Finally, a simple yet effective mechanism is established to correct the pseudo-labels by progressively fusing the outputs of the target model. In the second stage, the pseudo-label learning is discarded in favor of exploring the semantic structure. The cluster separation is designed to make the average outputs of different clusters orthogonal for realizing cluster transfer. Extensive experiments demonstrate the superiority of our method.