HIMO: Cross-Arbitrary-Modality Image Invariant Feature Transform With Hierarchical Intrinsic Major Orientation

Area of Science:

• Computer Vision
• Image Processing
• Pattern Recognition

Background:

• Invariant feature extraction is crucial for intelligent image processing with multi-source/modal imaging.
• Current cross-modal matching methods often focus on specific modalities, lacking a general approach.

Purpose of the Study:

• To develop a general approach for cross-arbitrary-modal image invariant feature extraction and matching.
• To introduce a handcrafted invariant feature transform inspired by human vision.

Main Methods:

• Proposed Hierarchical Intrinsic Major Orientation (HIMO), a non-data-driven algorithm based on Intrinsic Major Orientation (IMO) extraction.
• HIMO includes Difference-of-Feature Suppression (DoFS) keypoint detector, Polar-Pyramid descriptor (PolarP), and Cascaded Dynamic Multi-scale Strategy (CDMS).
• Introduced two large-scale cross-modal datasets (GCZ and WDS) and evaluation metrics.

Main Results:

• HIMO effectively handles intensity distortion, rotation, scale differences, geometric deformation, and noise.
• Comprehensive experiments on 5 datasets showed HIMO's superior performance against 10 traditional and 15 deep-learning methods.
• Demonstrated robustness, stability, and generalization across diverse imaging conditions.

Conclusions:

• HIMO offers a universal solution for cross-modal image matching.
• The proposed method achieves state-of-the-art performance in challenging cross-modal scenarios.
• HIMO provides a robust and generalizable feature extraction technique for intelligent image processing.