ADMD: An Accelerated Denoising Masked Diffusion Model for Unsupervised Anomaly Detection in Consumer Electronics Manufacturing

Due to the unnecessary of labeled samples and the benefit of low costs, unsupervised anomaly detection plays a vital role in guaranteeing the product qualities of consumer electronics manufacturing industry. Existing methods still suffer from limited feature extraction capability and substandard detection accuracy. As diffusion models (DMs) exhibit good image generating effects by approximating the normal samples’ distributions through denoising, DMs are thus widely applied in unsupervised anomaly detection. In this work, in order to enhance the DM’s denoising efficiency, we propose an Accelerated Denoising Masked Diffusion (ADMD) model by optimizing the non-Markovian-based denoising step. We also utilize area division mask operations to preserve pixels in normal areas and improve the image reconstruction quality. Moreover, to improve the effectiveness of feature extraction and the accuracy of anomaly detection, we adopt the domain adaptation method to extract valuable features. We leverage both image-level score and feature-level score to obtain a balanced anomaly score heatmap. Extensive experiment results on the MVTec-AD dataset display the superiority of our ADMD model, the image-level and pixel-level average detection precisions can achieve 99.2% and 67.3% respectively.