Harmonized Domain Enabled Alternate Search for Infrared and Visible Image Alignment

Area of Science:

• Computer Vision
• Image Processing
• Sensor Fusion

Background:

• Infrared and visible image alignment is crucial for multi-modal perception applications.
• Manual calibration methods are labor-intensive, lack portability, and struggle in dynamic environments.
• Existing methods face challenges in addressing spectral and environmental variations causing positional and scale discrepancies.

Purpose of the Study:

• To propose a novel harmonized representation-based method for accurate and adaptable infrared and visible image alignment.
• To overcome the limitations of existing manual calibration techniques.
• To enhance the performance of multi-modal perception applications in dynamic scenarios.

Main Methods:

• Developed an invertible translation process to create a harmonized representation domain for infrared and visible images.
• Designed a hierarchical framework for coarse-to-fine deformation correction using advanced perception and residual estimation.
• Employed an alternate correlation search mechanism for precise correspondence matching.

Main Results:

• The proposed method achieves high accuracy and scene adaptability in infrared and visible image alignment.
• Experimental results demonstrate superior performance compared to state-of-the-art methods.
• Introduced the first benchmark dataset with ground truth for evaluating misaligned infrared and visible images.

Conclusions:

• The harmonized representation-based approach effectively aligns infrared and visible images, enhancing multi-modal perception.
• The proposed method offers a robust and adaptable solution for real-world applications.
• The new benchmark facilitates further research and development in image alignment techniques.