Learning a Deep Ensemble Network With Band Importance for Hyperspectral Object Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Hyperspectral videos (HSVs) offer rich material identification capabilities for object tracking.
• Current hyperspectral trackers often rely on handcrafted features, limiting performance due to insufficient training data for deep learning.
• A significant performance gap exists in hyperspectral object tracking due to the underutilization of deep learning.

Purpose of the Study:

• To develop an end-to-end deep ensemble network (SEE-Net) for hyperspectral video object tracking.
• To address the challenge of limited training data for deep feature extraction in hyperspectral tracking.
• To improve tracking accuracy and adaptability by learning band importance.

Main Methods:

• Established a spectral self-expressive model to learn band correlations and derive band importance.
• Developed a spectral self-expressive module for efficient, non-iterative learning of nonlinear mappings to band importance.
• Utilized band importance to create false-color images for deep feature extraction and ensemble tracking results.

Main Results:

• SEE-Net demonstrates high computational efficiency and rapid adaptation to target appearance changes.
• The method effectively suppresses unreliable tracking by weighting false-color images based on learned importance.
• Extensive experiments confirm SEE-Net's superior performance compared to state-of-the-art hyperspectral tracking approaches.

Conclusions:

• SEE-Net provides an effective deep learning framework for hyperspectral video object tracking.
• The learned band importance significantly enhances feature representation and tracking robustness.
• The proposed approach offers a promising direction for advancing hyperspectral image analysis and tracking applications.