Dissecting and Reassembling Color Correction Algorithms for Image Stitching.

Area of Science:

• Computer Vision
• Image Processing
• Computational Photography

Background:

• Existing color correction methods lack a unified classification system.
• Evaluating color correction effectiveness in real-world applications like image stitching is challenging.

Purpose of the Study:

• Introduce a compositional framework for classifying color correction methods.
• Generate and evaluate novel color correction methods through combinatorial assembly of computational units.
• Assess method performance using perceptual image quality metrics, focusing on real-world applications.

Main Methods:

• Developed a compositional framework based on two main computational units.
• Dissected 15 existing algorithms and introduced 4 new computational units.
• Created ~100 novel color correction methods by reassembling units combinatorially.
• Tested methods on diverse datasets, including real and artificial color transformations and a novel dataset for specific acquisition setups.
• Evaluated performance using recent perceptual image quality metrics.

Main Results:

• Combinations of new computational units excel in real-world image stitching scenarios, robust to misalignments and light scattering.
• For accurate alignment and artificial color alterations, methods using new units or novel combinations of existing units perform best.
• First-time experimental evidence provided using recent perceptual image quality metrics.

Conclusions:

• The proposed compositional framework effectively classifies and generates novel color correction methods.
• New computational units and their combinations offer significant improvements for challenging real-world image processing tasks like stitching.
• The study provides a robust evaluation methodology for color correction algorithms using advanced perceptual metrics.