Prompt-Guided Mamba Network for Multi-Class Anomaly Detection and Localization

现有的无监督异常检测方法在多类训练时往往受到类别间干扰的困扰，导致在实际应用中检测性能显著下降。为了解决这个问题，本文提出了一种新颖的基于提示引导的Mamba网络(PGM)用于多类异常检测和定位。首先，通过采用UniRepLKNet作为预训练特征提取器，模型有效压缩并记忆正常图像的大量信息。接着，基于Mamba结构的解码器结合了全局和局部建模能力能够有效重建多尺度特征。最后，PGM采用了分层类感知提示模块，动态地将类别特定信息编码到类别先验池中，以减轻异常类别之间的干扰。在MVTec AD数据集上进行的实验表明，所提出的PGM方法与大多数先进方法相比具有最佳性能，在图像级AUROC方面达到98.7%、像素级AUROC方面达到98.2%。同时，详尽的消融研究验证了模型每个组件的贡献。总体而言，所提出的PGM方法在异常检测领域具有显著优势。Existing unsupervised anomaly detection methods often suffer from significant performance degradation in practical applications due to inter-class interference during multi-class training. To address this issue, this paper proposes a novel Prompt-Guided Mamba Network (PGM) for multi-class anomaly detection and localization. Firstly, by employing UniRepLKNet as the pre-trained feature extractor, the model effectively compresses and memorizes a large amount of information from normal images. Subsequently, the decoder based on the Mamba structure combines global and local modeling capabilities to effectively reconstruct multi-scale features. Finally, PGM incorporates a hierarchical class-aware prompt module that dynamically encodes class-specific information into a class prior pool to mitigate interference between different anomaly classes. Experiments conducted on the MVTec AD dataset demonstrate that the proposed PGM method outperforms most state-of-the-art methods, achieving 98.7% in image-level AUROC and 98.2% in pixel-level AUROC. Additionally, extensive ablation studies validate the contribution of each component of the model. Overall, the proposed PGM method shows significant advantages in the field of anomaly detection.