RDDPM: Robust Denoising Diffusion Probabilistic Models for Unsupervised Anomaly Detection and Segmentation in Contaminated Data

Abstract

Anomaly detection in noisy images is essential for quality control in industries like steel, composites, and textiles. Traditional unsupervised anomaly segmentation models, including Robust Principal Component Analysis and Smooth Sparse Decomposition, rely on restrictive data assumptions including anomaly sparsity and a low-rank structure in the normal background. On the other hand, while diffusion-based anomaly detection methods are effective in handling non-linear patterns and non-sparse anomalies, they require large amounts of normal images for training, limiting their applicability in contaminated data. In this paper, we propose a novel method called Robust Diffusion, which leverages robust regression to train a diffusion model without uncontaminated data. Through simulations and a real-world case study on solar cells, we demonstrate that our approach outperforms both traditional statistical models and existing diffusion-based methods in detection accuracy. We also introduce a benchmark synthetic dataset, the RDDPM dataset, containing 100,000 normal and 50,000 anomalous samples with complex patterns and diverse anomaly types under varying lighting intensities. This dataset is designed to rigorously evaluate models for unsupervised anomaly segmentation in contaminated data with complex patterns.