Cross-Spectral Local Descriptors via Quadruplet Network

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Cross-spectral image matching is crucial for tasks like surveillance and remote sensing.
• Existing methods often struggle with the domain shift between different spectral bands.
• Triplet networks show promise but require adaptation for cross-spectral challenges.

Purpose of the Study:

• To develop a novel Convolutional Neural Network (CNN) architecture, Q-Net, for learning robust local feature descriptors.
• To enable effective image patch matching across different spectral bands (e.g., visible and near-infrared).
• To improve upon the state-of-the-art in cross-spectral matching performance.

Main Methods:

• A quadruplet network architecture (Q-Net) is proposed, trained on matched and non-matching cross-spectral image pairs.
• Image patches are mapped to a common Euclidean space, invariant to the input spectral band.
• The approach adapts successful triplet network concepts for cross-spectral scenarios, addressing unique non-matching pair challenges.

Main Results:

• Q-Net significantly improves the state-of-the-art performance on a public cross-spectral VIS-NIR dataset.
• The method demonstrates effectiveness in learning local feature descriptors for cross-spectral matching.
• The technique achieves comparable performance to triplet networks in mono-spectral settings with reduced training data.

Conclusions:

• The proposed Q-Net architecture offers a powerful solution for cross-spectral image patch matching.
• Q-Net provides improved accuracy and efficiency compared to existing methods.
• The adaptability of Q-Net to mono-spectral tasks highlights its versatility.