Multi-scale feature-enhanced Auto-UNet with Cross Graph Mamba for high resolution image anomaly detection

High resolution image provides the granular detail necessary for precise anomaly detection, However, limited anomalous samples and intensive computation challenge the high resolution image anomaly detection performance. While recent advances in Autoencoders, Graph Neural Networks, and Mamba have shown promise, they face critical bottlenecks: (1) existing methods often suffer from complex training regimes and suboptimal accuracy; (2) Graph Neural Networks are prone easily to overfitting and over-smoothing; (3) Mamba models exhibit high sensitivity to input sequence ordering. To address these challenges, we propose a Multi-Scale Image and Graph Feature Enhanced Auto-UNet with Cross Graph Mamba. The 2D Discrete Wavelet Transform at different scales with their corresponding inverse transforms serve as the parameter-efficient Auto-UNet. The lightweight architecture of Mamba further improves the computational efficiency. We perform feature enhancement in both the wavelet and spatial domain, adopting the Butterworth high pass operator, graph correlation and cross-correlation, respectively. Then, Cross Graph Mamba strengthens the detection accuracy of anomalous regions via cross message propagation between graph correlation and graph cross-correlation feature enhancements. Furthermore, we leverage Principal Component Analysis prompt tuning and Normalized Simulated Annealing algorithm within Cross Graph Mamba to address issues caused by imbalanced samples. Regarding Mamba’s sensitivity to sequences, we design an Information Entropy Sorting mechanism that dynamically ranks nodes. Finally, we utilize sum of Mahalanobis distances and weighted focal loss as loss function. Extensive experiments on four benchmark datasets demonstrate that our method significantly outperforms SOTAs in both detection accuracy and computational efficiency.