.png){kind=icon}
.png){kind=icon}
Return
PixDiff: Multiresolution Diffusion Network With Pixelization for Hyperspectral Anomaly Detection
.png){kind=icon}
PRE
AI
DOI：10.1109/TGRS.2026.3657039.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Hyperspectral anomaly detection (HAD) is a crucial area in hyperspectral image (HSI) processing, aiming to identify unknown anomalous objects. Since anomalies are usually more difficult for networks to learn compared with backgrounds, a popular deep learning-based paradigm is to distinguish anomalies from the background through reconstruction errors. However, anomaly regions in the reconstruction objective may still guide the network during training. In addition, complex backgrounds in HSI that are difficult to reconstruct may be challenging to differentiate from anomalies. Due to the unique spatial characteristics of anomalies and complex backgrounds, structural spatial transformations can be applied to create distinguishable differences between them. In this article, a multiresolution diffusion network based on pixelization called PixDiff is proposed to reduce the visibility of anomalies and differentiate them from complex backgrounds. In PixDiff, multiscale pixelization is innovatively introduced into HAD to preserve coarse background structures while disrupting anomaly information. Then, a structural isolation module (SIM) is proposed to distinguish difficult-to-reconstruct complex backgrounds from potential anomalies through interresolution differences. Furthermore, a multiresolution adaptive weighted reconstruction loss is proposed to further increase the distinction between anomalies and complex backgrounds. Experimental results on five HAD datasets demonstrate that the proposed PixDiff achieves superior detection performance compared with existing state-of-the-art methods.
Keywords:
Anomaly detection
diffusion model
hyperspectral image (HSI)
multiresolution network
pixelization
Journal
IF:
8.6
Papers:
2.1W
Citations:
10.7W