Fusion of label attention and mask attention for unsupervised and self-supervised anomaly detection

Abstract

Many uncontrollable factors during industrial flow lead to products’ unforeseen defects. Anomaly detection for the defects is inclined to use unsupervised frameworks to formulate one-class classification tasks, as well as segmentation tasks. The attention mechanism has proved to be effective in defect classification and segmentation. However, previous works with attention mechanisms have not investigated the combination of the two kinds of attention for the above tasks in depth. In this paper, we propose a framework that fuses Label Attention and Mask Attention (LAMA). Specifically, the LAMA is implemented by a loss prediction module and a decoder in a transform-based segmentation model. Moreover, a network based on normalizing flow is introduced as the architecture of anomaly detection, which integrates with the LAMA. The experimental results on MVTecAD demonstrate that the proposed method outperforms the contrast algorithms. The proposed framework achieves an overall image-wise AUC of 99.6% and 98.6% on pixel-wise AUC.