Agm-Net: Attention-guided masking denoising anomaly location network

Unsupervised anomaly detection based on knowledge distillation (KD) has been empirically verified to be effective. Nevertheless, previous KD models were constrained in terms of performance and generalization ability due to the similarity of the architectures of the student network (S) and the teacher network (T). Hence, this paper proposes an attention-guided masked denoising anomaly localization network (Agm-Net). Firstly, an attention-guided U-shaped denoising architecture was incorporated into the student network, which enhanced the structural difference of the T-S model while guiding the denoising process. Secondly, a feature-level mask generation module was introduced, adopting a strategy of regionally random masking to make the size of each masked region controllable. This prevented the loss of feature details and excessive recovery time caused by excessive masking and strengthened the recovery ability of local information. Finally, a randomly connected boundary smoothing anomaly synthesis strategy was proposed to synthesize defect images with diverse shapes that closely resemble real anomalies, enhancing the model's understanding of anomalous data. The Agm-Net model was compared and evaluated with existing advanced methods on anomaly detection datasets MVTec AD, VisA, and the real PCB sample dataset BHAAD. The pixel-level AU-ROC and PRO reached 98.2 % and 94.6 % on the MVTec AD dataset, 99.2 % and 95.1 % on the VisA dataset, and 98.9 % and 93.4 % on the real PCB dataset BHAAD, respectively. This validates the effectiveness of the Agm-Net.