DFP-AD: A Dynamic Frequency Prototype Framework for Unsupervised Anomaly Detection

doi:10.1016/j.inffus.2026.104728

DFP-AD: A Dynamic Frequency Prototype Framework for Unsupervised Anomaly Detection

2026-08-22

0

PRE

AI

DOI：10.1016/j.inffus.2026.104728

Original

Original request for help

Save

Abstract

En 中文

• Learns adaptive frequency prototypes for unsupervised anomaly detection. • Disentangles high- and low-frequency cues for better structural modeling. • Correlation coherence loss improves feature-prototype alignment. • Cross-frequency modulation enables adaptive multi-scale reconstruction. • Achieves robust performance across multi-, single-class, and few-shot tasks. Abstract Unsupervised anomaly detection aims to identify abnormal patterns without labeled anomalies while generalizing across diverse settings. Prototype-based approaches address this by learning normal feature centers and detecting deviations as anomalies. However, existing methods predominantly operate in the spatial domain, where appearance-level representations entangle structural information with texture variations and imaging conditions, causing a mismatch between the modeled normality and the nature of anomalies. Consequently, (1) the lack of frequency-aware modeling limits the capture of structural regularities and subtle spectral disruptions, and (2) spatial prototype-guided decoding imposes static similarity constraints without adaptive multi-scale modulation, hindering accurate reconstruction and generalization across diverse anomaly patterns. To address these challenges, we propose DFP-AD, a dynamic frequency prototype-based approach that learns adaptive frequency prototypes and modulates reconstruction via cross-frequency interactions. Specifically, to capture frequency-specific cues, we design an adaptive frequency prototype extractor that dynamically demodulates features and disentangles high- and low-frequency representations, enabling more discriminative prototype learning. To further enhance representation alignment, we devise a correlation coherence loss that encourages strong linear correlations between features and their most relevant prototypes. Furthermore, to improve semantic adaptability during decoding, we introduce a prototype-aligned frequency modulator that dynamically aligns decoded features with high- and low-frequency prototypes, assigning frequency-specific weights to refine anomaly representations. These components jointly enable DFP-AD to achieve robust and accurate anomaly localization. Extensive experiments demonstrate competitive image- and pixel-level performance across diverse unsupervised anomaly detection settings. The code is available: https://github.com/zhangqiao970914/DFP-AD

Keywords:

Unsupervised anomaly detection

Prototype learning

Frequency modeling

DFP-AD: A Dynamic Frequency Prototype Framework for Unsupervised Anomaly Detection

Abstract

Journal

Organization

Cited Papers