MambaAD: Exploring State Space Models for Multi-class Unsupervised Anomaly Detection

MambaAD: Exploring State Space Models for Multi-class Unsupervised Anomaly Detection

14 Apr 2024 | Haoyang He1*, Yuhu Bai1*, Jiangning Zhang2, Qingdong He2, Hongxu Chen1, Zhenye Gan2, Chengjie Wang2, Xiangtai Li3, Guanzhong Tian1, Lei Xie1
The paper introduces MambaAD, a novel approach for multi-class unsupervised anomaly detection using the Mamba framework. MambaAD combines a pre-trained encoder and a Mamba-based decoder with Locality-Enhanced State Space (LSS) modules at multiple scales. The LSS module integrates Hybrid State Space (HSS) blocks and parallel multi-kernel convolutions to capture both global and local information. The HSS block uses Hybrid Scanning (HS) encoders to encode feature maps into five scanning methods and eight directions, enhancing global connections through the State Space Model (SSM). The method demonstrates superior performance on six diverse anomaly detection datasets and seven metrics, achieving state-of-the-art results while maintaining low computational complexity and model parameters. The code and models are available at <https://lewandofskie.github.io/projects/MambaAD>.The paper introduces MambaAD, a novel approach for multi-class unsupervised anomaly detection using the Mamba framework. MambaAD combines a pre-trained encoder and a Mamba-based decoder with Locality-Enhanced State Space (LSS) modules at multiple scales. The LSS module integrates Hybrid State Space (HSS) blocks and parallel multi-kernel convolutions to capture both global and local information. The HSS block uses Hybrid Scanning (HS) encoders to encode feature maps into five scanning methods and eight directions, enhancing global connections through the State Space Model (SSM). The method demonstrates superior performance on six diverse anomaly detection datasets and seven metrics, achieving state-of-the-art results while maintaining low computational complexity and model parameters. The code and models are available at <https://lewandofskie.github.io/projects/MambaAD>.
Reach us at info@study.space