Sub-pixel Image Registration via Continuous Domain Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image registration methods are limited in achieving sub-pixel accuracy, which can result in poorer quality image stabilization and noticeable jitter in video applications, especially when tracking moving features.
Innovation Solution
A method for sub-pixel image registration that computes quantities from pixel values of a template and target images, optimizing an objective function defined on a continuous domain to determine a globally optimal registration location, allowing for accurate and efficient tracking of templates in video frames.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If pixel-accurate registration is used, then computational complexity is reduced, but registration precision deteriorates
Solution Approach 1:
The patent segments the continuous registration space into discrete pixel locations and fractional pixel offsets. By separating the integer pixel search from the fractional offset optimization, the method efficiently explores the search space without requiring exhaustive computation at every possible sub-pixel location, thus achieving high precision without proportional computational complexity increase.
Solution Approach 2:
The patent performs preliminary coarse registration at pixel accuracy to establish a base location, then refines this result with sub-pixel precision. This preliminary action eliminates the need to perform computationally expensive sub-pixel calculations across the entire image, concentrating computational effort only where needed around the preliminary match location.
2Measurement precision
If iterative sub-pixel registration is used, then registration precision is improved, but computational time increases
Solution Approach 1:
The patent performs a preliminary pixel-accurate registration to establish a starting point, then uses this as the basis for limited iterative refinement. This preliminary action significantly reduces the number of iterations needed for convergence compared to starting iterative refinement from scratch, thereby reducing total computational time while maintaining sub-pixel precision.
Solution Approach 2:
The patent applies partial iteration by performing only the necessary number of refinement iterations around the preliminary match location rather than exhaustive iteration across the entire search space. This partial action achieves sufficient sub-pixel precision without the excessive computational time that would result from complete exhaustive search.
3Productivity
If template search region is restricted to near previous location, then processing speed is improved, but tracking accuracy deteriorates
Solution Approach 1:
The patent uses the previous template location as a preliminary estimate to define a restricted search region, then performs sub-pixel refinement within this region. This preliminary action based on temporal coherence maintains processing speed by limiting the search space, while the sub-pixel refinement ensures tracking accuracy is not compromised by the restriction.
Solution Approach 2:
The patent changes the precision parameter from integer pixels to fractional pixels within the restricted search region. By maintaining the same spatial search region boundaries but increasing the precision of location measurement and optimization, the method achieves higher tracking accuracy without expanding the search region or sacrificing processing speed.
Data Source
AI summary
An approach to sub-pixel image registration involves determining parameters of an objective function from pixel values of a template image and of a target image. This objective function is defined on a bounded continuous domain of relative displacements of the template image and the target image, which corresponds to sub-pixel registration locations. This objective function is directly optimized without necessarily evaluating spatially interpolated values of either the template image or the target image to achieve a global optimum of the objective function in the bounded domain that provides the sub-pixel registration location. The approach can be used for tracking a template in a sequence of video frames.