MASF: MASKED ASYMMETRIC SPECTRAL FLOW FOR UNSUPERVISED INDUSTRIAL ANOMALY DETECTION

Abstract

An important problem in the industrial quality control application is that of unsupervised anomaly detection, in which only defect free training images are available. All the currently available state-of-the-art techniques such as PaDiM, PatchCore, RD++, CFA, ISSTAD, and ADTR are in the spatial domain and miss the part of the information in the frequency domain of CNN feature maps that has anomaly-discriminative information. We introduce the first framework to leverage frequency domain at the CNN feature level, called MASF (Masked Asymmetric Spectral Flow). MASF introduces five novel components: (1) a Spectral Frequency Decomposition Module (SDM) based on 2D FFT on intermediate feature maps; (2) Asymmetric Masked Feature Distillation (AMFD) using dual spatial-frequency domain masking and Spectral-Spatial Cross-Attention (SSCA) fusion; (3) a Spectral-Anchored Memory Bank (SAMB) for rotation-robust prototype retrieval; (4) Uncertainty-Gated Hierarchical Score Fusion (UGHF) with learnable per-scale precision weights; and (5) Test-Time Spectral Augmentation (TTSA) by FFT phase perturbation. On the Bottle category, evaluated on the MVTec Anomaly Detection benchmark, MASF gets Image-AUROC = 99.9999, Pixel-AUROC = 0.9853, PRO = 0.9461, and AP = 99.9999. In 11 stable training categories, MASF reaches the performance of mean Image-AUROC = 0.8008 and mean Pixel-AUROC = 0.9380 with just 15 training epochs. The design of MASF is directly motivated by the results of FFT spectral analysis which shows that the frequency signature of the normal image is different from that of the anomalous image, as shown by industrial images.