ShufflePaste: Four-Way Structural Disturbance Learning With Swin Transformer for Industrial Anomaly Detection

Abstract

Industrial anomaly detection aims to identify defective samples that deviate from normal visual patterns, while real anomaly annotations are often scarce, costly, and incomplete in practical manufacturing scenarios. To address this challenge, this paper proposes ShufflePaste, a self-supervised representation learning framework for one-class industrial anomaly detection based on four-way structural disturbance learning and a Swin Transformer backbone. Unlike conventional CutPaste-based methods that mainly simulate appearance discontinuities by pasting local patches, the proposed Local Shuffle strategy alters the local spatial arrangement by randomly permuting grid cells within a selected region. Together with normal samples, CutPaste-Normal samples, and CutPaste-Scar samples, Local Shuffle forms a four-way pretext task that encourages the model to learn structure-sensitive representations. A Swin Transformer is further adopted to capture both local and long-range structural dependencies through hierarchical shifted-window attention. During inference, image-level anomaly scores are computed from Swin-T backbone features using Gaussian density estimation and the Mahalanobis distance; the projection head and classifier outputs do not contribute to anomaly scoring. Experiments on the MVTec AD dataset yield an average image-level ROC-AUC of 98.3% and 100.0% ROC-AUC on several categories. These results indicate the potential of combining local structural disturbance learning with Transformer-based representation modeling for industrial anomaly detection.