Parallax-Tolerant Weakly-Supervised Pixel-Wise Deep Color Correction for Image Stitching of Pinhole Camera Arrays.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Image stitching commonly uses algorithms to create panoramas, but faces artifacts like parallax and color shifts due to varying viewpoints.
• Existing solutions struggle with parallax and general color differences, especially in pinhole camera arrays.

Purpose of the Study:

• To develop a robust framework for parallax-tolerant, weakly supervised, pixel-wise deep color correction in image stitching.
• To address limitations of current methods in handling color discrepancies and parallax in panoramic images.

Main Methods:

• A two-stage framework employing a convolutional module for feature extraction and a parallax-tolerant color correction network with dynamic loss weights.
• Utilizing a gradient-based Markov Random Field for harmonizing non-overlapping regions with overlapping ones.
• Introducing a novel evaluation metric, Color Differences Across the Seam, for quantitative assessment.

Main Results:

• The proposed framework effectively compensates for color differences in overlapping regions using adaptive techniques.
• Non-overlapping regions are harmonized with overlapping regions, ensuring visual consistency.
• Demonstrated superior performance over existing methods in both qualitative and quantitative evaluations on benchmark datasets and real-world images.

Conclusions:

• The developed framework successfully eliminates visible artifacts in stitched images from pinhole camera arrays.
• The approach produces natural-looking composite images by effectively managing parallax and color differences.
• The new evaluation metric provides a reliable measure for assessing the quality of image stitching transitions.