1. Multi-scale frequency-guided two-stream network for hyperspectral anomaly detection
- Author
-
Zhe Zhao, Jiangluqi Song, Dong Zhao, Jiajia Zhang, Huixin Zhou, and Jun Zhou
- Subjects
Hyperspectral image (HSI) ,Anomaly detection ,Frequency-decomposition ,Two-stream network ,Physical geography ,GB3-5030 ,Environmental sciences ,GE1-350 - Abstract
Hyperspectral anomaly detection (HAD) holds significant importance in remote sensing image processing and has been recently empowered by deep learning-based methods. Despite achieving good performance by reconstructing the background and suppressing abnormal targets, these methods often fail to distinguish the attributes of latent features during background reconstruction, resulting in unsatisfactory reconstruction quality. This occurs because the background usually has low-frequency properties and the anomalies have high-frequency properties, so these issues should be considered during background reconstruction. To overcome this weakness, a multi-scale frequency-guided two-stream network (MFTNet) is proposed for HAD. Firstly, considering the different frequency attributes of background and anomalies, we use an encoder to extract features and then design a multi-scale frequency decomposition module to decompose the features into high-frequency (HF) part for anomalies and the low-frequency (LF) part for background. Furthermore, different from anomaly suppression-based methods, we introduce a high-frequency enhancement module to highlight abnormal targets in the HF part, making them stand out in the background. Meanwhile, the features from the encoder contain more background information, so we fuse the encoded features and LF part features by a self-attention module to accurately recover the background. Finally, we perform anomaly detection on the reconstructed anomaly and background parts, respectively, and the average of those two results is the final detection map. Comparative analysis on six datasets against thirteen baseline methods, including seven effective deep learning-based methods, demonstrates that our proposed MFTNet delivers competitive detection results. The code of this work will be released at: https://github.com/xautzhaozhe/MFTNet.
- Published
- 2025
- Full Text
- View/download PDF