Learning Disentangled Priors for Hyperspectral Anomaly Detection: A Coupling Model-Driven and Data-Driven Paradigm.

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Hyperspectral anomaly detection (HAD) faces challenges due to inadequate prior knowledge modeling.
• This limitation creates a performance bottleneck in accurately distinguishing background from anomalous objects.

Purpose of the Study:

• To develop a novel hyperspectral anomaly detection (HAD) method by integrating model-driven and data-driven techniques.
• To enhance the modeling of prior knowledge for improved background representation and anomaly extraction.

Main Methods:

• Introduced a learning disentangled priors (LDP) paradigm, coupling low-rank representation (LRR) with deep learning.
• Employed a model-driven deep unfolding architecture separating explicit (low-rank) and implicit (deep network) priors.
• Utilized a skip residual connection to model interdependencies between explicit and implicit priors.

Main Results:

• The proposed LDP method demonstrated superior performance compared to existing advanced HAD techniques.
• Experiments on multiple datasets confirmed LDP's enhanced detection accuracy and generalization capabilities.
• Mathematical convergence proof was provided for the LDP model.

Conclusions:

• The LDP paradigm effectively addresses the prior knowledge modeling challenge in HAD.
• This approach offers a significant advancement in hyperspectral image analysis for anomaly detection.
• LDP shows strong potential for real-world applications requiring accurate hyperspectral data interpretation.