Learning Tensor Low-Rank Representation for Hyperspectral Anomaly Detection.

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Low-rank representation (LRR) methods are common for hyperspectral anomaly detection.
• Existing LRR models convert 3D hyperspectral images (HSIs) to 2D matrices, losing crucial 3D structural information.
• This loss of structural information can hinder effective background and anomaly separation.

Purpose of the Study:

• Propose a novel tensor low-rank and sparse representation (TLRSR) method for hyperspectral anomaly detection.
• Preserve the intrinsic 3D structure of HSIs during anomaly detection.
• Improve the separation of background and anomalous components in HSIs.

Main Methods:

• Developed a 3D tensor low-rank model to separate the background, represented by a tensorial background dictionary and coefficients.
• Utilized weighted tensor nuclear norm and L_F,1 sparse norm for dictionary design, enhancing background relevance.
• Incorporated Principal Component Analysis (PCA) as a preprocessing step to reduce computational load while retaining HSI object information.
• Employed Alternating Direction Method of Multipliers (ADMMs) for efficient model solving.

Main Results:

• The proposed TLRSR method effectively separates background and anomalies by leveraging the 3D structure of HSIs.
• Experimental comparisons demonstrate the competitiveness of TLRSR against state-of-the-art hyperspectral anomaly detection algorithms.
• The use of PCA preprocessing reduces computational time without significant loss of object information.

Conclusions:

• The novel TLRSR method offers a significant advancement in hyperspectral anomaly detection by preserving 3D structural properties.
• TLRSR provides superior performance in separating complex backgrounds and identifying anomalies compared to existing 2D-based LRR methods.
• The method is computationally efficient and demonstrates state-of-the-art results in hyperspectral anomaly detection tasks.