Continuous Diffusion Models with Explicit Score Matching for Highly Efficient Anomaly Detection

Fengyu Liang, Guowen Kuang, Xu Wang, Jinfeng Yang, Peiyu Guan

Keywords: Diffusion models, Explicit score matching, variance-guided, Anomaly detection

TL;DR: We propose a diffusion model framework based on explicit score matching and a variance-guided process for anomaly detection, achieving SOTA performance with high efficiency.

Abstract: Diffusion models have proven to be highly effective in generating high-quality reconstructed images, making them ideal for the rigorous requirements of reconstruction-based anomaly detection systems. While continuous diffusion mod- els unify discrete implementations, existing methods predom- inantly rely on denoising score matching (DSM), as directly acquiring explicit scores remains challenging. In this study, we introduce a novel diffusion framework utilizing explicit score matching (ESM) via a dual-stream neural network, trained by maximum likelihood estimation. Based on the first systematic comparison between DSM and ESM paradigms, variance-guided diffusion process is developed to further im- prove the performance. Comprehensive experimental evalua- tions confirm the superior anomaly detection capabilities and computational efficiency of the proposed system. The frame- work’s flexibility allows seamless integration with existing diffusion models, offering a potential pathway for broader ap- plications in generative tasks.

Primary Area: generative models

Submission Number: 7498