LFQUIAD: Lookup-Free Quantized autoencoder for few-shot Unsupervised Industrial Anomaly Detection via Synthetic Diffusion Inpainting
Keywords: anomaly detection, Synthetic Diffusion Inpainting
TL;DR: We propose a lookup-free quantized autoencoder (LFQ-AE) for medical anomaly detection that leverages diffusion-based inpainting to reconstruct healthy regions and identify abnormal patterns, achieving high accuracy without codebook overhead.
Abstract: Unsupervised anomaly detection (UAD) is crucial in industrial and medical applications, offering scalable and cost-efficient alternatives to manual inspection by detecting abnormal patterns without requiring labeled anomalies. However, real-world anomalies are often scarce and ambiguous, limiting the effectiveness of conventional methods. We propose \textbf{LFQUIAD}, a novel UAD framework that integrates a quantization-driven autoencoder with a modular \textbf{Anomaly Generation Module (AGM)}. AGM generates diverse and semantically meaningful synthetic anomalies using prompt-guided, diffusion-based inpainting, providing pixel-level supervision in few-shot scenarios. This enables robust model training without real anomaly data. At the core of LFQUIAD lies \textbf{Lookup-Free Quantization (LFQ)}, a codebook-free representation learning method that discretizes features with high precision while improving generalization and robustness.Our method achieves state-of-the-art performance on MVTecAD and VisA benchmarks, excelling in both anomaly detection and segmentation under limited-data conditions. The plug-and-play nature of AGM also allows seamless integration into other detection pipelines, making LFQUIAD a practical and effective solution for real-world anomaly detection tasks.
Submission Number: 71
Loading