Feature Interaction Learning Network for Cross-Spectral Image Patch Matching.

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Existing feature relation learning methods for cross-spectral image patch matching extract shallow features, leading to feature loss or noise.
• Current multi-branch networks are parameter-heavy, despite better feature extraction.

Purpose of the Study:

• To propose a novel two-branch feature interaction learning network (FIL-Net) for effective cross-spectral image patch matching.
• To develop a new feature interaction learning module for mining common and private features.
• To create stronger feature extraction networks and a multi-loss optimization strategy.

Main Methods:

• Introduced a novel feature interaction learning idea and module to mine common/private features between cross-spectral patches.
• Developed a two-branch residual feature extraction network for enhanced feature extraction capabilities.
• Proposed a multi-loss strong-constrained optimization strategy for efficient, invariant, and discriminative feature extraction.

Main Results:

• FIL-Net effectively mines common and private features, extracting richer, deeper feature relations with invariance and discriminability.
• The two-branch residual network demonstrates stronger feature extraction capabilities.
• Extensive experiments confirm FIL-Net achieves state-of-the-art performance across three cross-spectral image patch matching scenarios.

Conclusions:

• The proposed FIL-Net offers an efficient and effective solution for cross-spectral image patch matching.
• The novel feature interaction learning and optimization strategy significantly improve feature relation extraction.
• FIL-Net advances the field by providing state-of-the-art performance and new benchmark datasets.