Tile Image Stitching via NCC and Weighted Least Squares
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Feature-based image registration methods often fail to accurately merge images due to registration errors, especially when the number of feature points is limited, leading to suboptimal results in applications like panorama imaging and medical imaging.
Innovation Solution
A computer-implemented method that performs local registration using normalized cross-correlation scores within a multi-resolution framework to determine spatial relationships between adjacent greyscale tile images, followed by global placement using weighted least squares, ensuring accurate alignment and stitching of images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If feature-based image registration is used to merge images, then the registration process can be computationally efficient, but registration accuracy deteriorates due to limited feature points
Solution Approach 1:
The patent changes the fundamental parameter used for registration from feature-based matching to pixel-based normalized cross-correlation (NCC) scoring. This allows the system to utilize all pixels in overlapping regions rather than relying on a limited set of extracted features, thereby improving registration accuracy while maintaining computational feasibility through the NCC metric
Solution Approach 2:
The patent creates multiple copies of the image at different resolutions (multi-resolution framework) to enable efficient computation of NCC scores across all pixels. By working with downsampled versions for initial matching and progressively refining with higher resolutions, the system achieves full-pixel utilization without prohibitive computational cost
2Measurement precision
If all pixels in overlapping regions are utilized for registration, then registration accuracy improves, but computational cost increases
Solution Approach 1:
The patent segments the computation process into multiple resolution levels. It divides the full-resolution image into progressively downsampled versions, performing NCC scoring at each level. This segmentation allows the system to process all pixels efficiently by distributing the computational load across multiple coarser representations before refining at full resolution
Solution Approach 2:
The patent performs preliminary registration actions at lower resolutions before finalizing at full resolution. By computing candidate matches using NCC on downsampled images first, the system pre-establishes potential correspondences that guide the subsequent full-resolution refinement, reducing the overall computational burden while maintaining accuracy
3Measurement precision
If local registration is performed on each adjacent tile pair, then spatial relationships are accurately determined, but the overall stitching process becomes complex
Solution Approach 1:
The patent merges the local registration results from all adjacent tile pairs into a unified global placement solution using weighted least squares. By combining individual NCC-based spatial relationships with a global optimization framework, the system resolves inconsistencies and achieves coherent overall stitching while maintaining the accuracy benefits of local pairwise registration
Data Source
AI summary
Techniques are provided for stitching tile images. A local registration can be performed to determine spatial relationships between pairs of adjacent tile images, in which the spatial relationships may be determined using normalized cross correlation (NCC) scores computed within a multi-resolution framework. A global placement can be performed to position all scanned tile images relative to one another. The global placement is determined with weighted least squares utilizing the determined spatial relationships between all adjacent greyscale tile images and the NCC scores as weights.