Manifold-Constrained Dynamic Decoupling Learning for Unsupervised Multiclass Anomaly Detection

While current unsupervised multiclass anomaly detection methods aim to build unified models for industrial applications, they face a critical dilemma between generalization capability and localization precision. Existing approaches using fixed encoders risk anomalous feature contamination during reconstruction, whereas adaptive encoders sacrifice cross-category generalization through single-class overfitting. To address this fundamental contradiction, we present manifold-constrained dynamic decoupling (MCDD) learning for unsupervised multiclass anomaly detection, which achieves dual constraints on normal feature manifolds through refinement of multiscale features from frozen encoders and robust reconstruction with learnable decoders. Specifically, we first propose the cross-hierarchy attentive bottleneck (CHAB) module, employing channel–spatial dual-domain attention gating to filter shallow texture features and deep structural features, constructing hybrid-scale normal base features. Furthermore, the noise-augmented feature expansion (NAFE) module locates critical encoder regions through attention mechanisms and injects learnable Gaussian noise during decoder upsampling, forcing reconstruction to focus on essential normal attributes. In addition, we construct the hybrid perception reasoning decoder (HPR-Decoder), integrating Visual Mamba’s long-range dependency modeling with graph attention convolution’s local correlation reasoning to achieve fine-grained generation of pixelwise anomaly maps. Experiments on MVTec AD and VisA datasets demonstrate that our method maintains superior multiclass detection performance with a single model while keeping model parameters within a reasonable range.