Image Registration via Pyramid Sub-sampling and Directional Motion Vectors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image-registration methods struggle with efficiently registering images with very different properties, such as those captured under varying exposure times or using different modalities, often losing edge information and requiring excessive computation.
Innovation Solution
The method employs image pyramids generated through sub-sampling, determining optimal movement directions from five possible directions, and updating motion vectors for each level of the pyramids using a probability-based method, reducing computation and stabilizing image registration across diverse image properties.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional probability-based methods (mutual information) are used for image registration, then registration stability is improved for images with very different properties, but computation time increases significantly
Solution Approach 1:
The patent divides the image into multiple resolution levels using image pyramids. Instead of processing the full-resolution image directly, the method segments the computation across multiple scales (e.g., 4 levels), calculating objective functions at each level. This segmentation reduces the computational burden at each stage while maintaining registration accuracy through progressive refinement from coarse to fine levels.
Solution Approach 2:
The patent performs preliminary coarse registration at lower resolution levels before refining at higher levels. By first establishing approximate alignment at coarser scales (where computation is cheaper), the method prepares initial motion vectors that guide subsequent finer-level registration. This preliminary action at reduced complexity prevents the need for exhaustive computation at full resolution.
2Adaptability or versatility
If image registration is performed on images with very different properties (different exposure times, modalities), then comprehensive image matching is achieved, but edge information is lost and motion estimation fails
Solution Approach 1:
The patent transitions from single-scale spatial processing to multi-scale spatial-frequency processing by constructing image pyramids. Each level of the pyramid represents a different spatial scale, allowing the system to capture both coarse structural information (at higher levels) and fine edge details (at lower levels). This dimensional change in processing scale enables robust matching across different image modalities and exposure conditions.
Solution Approach 2:
The patent changes the resolution parameter systematically across different levels of the image pyramid. By processing images at multiple resolution parameters (from coarse to fine), the method adapts to different image properties at appropriate scales. Coarse levels handle global alignment robustly, while fine levels recover edge information, thus preventing information loss that would occur at any single scale.
3Measurement precision
If full-resolution images are processed for accurate registration, then registration precision is improved, but computation time increases excessively
Solution Approach 1:
The patent segments the full-resolution processing task into multiple resolution levels. Computation is distributed across levels, with each level processing a subset of the total detail. This segmentation allows the system to achieve accurate registration through cumulative refinement rather than requiring all computational resources at full resolution simultaneously, thus improving processing speed while maintaining precision.
Solution Approach 2:
The patent performs preliminary registration operations at reduced resolution levels before finalizing at full resolution. Initial motion estimation and alignment are established at coarser scales where computation is faster, providing a head start that reduces the computational burden required for final precision registration. This preliminary action at lower computational cost accelerates the overall process.
Data Source
AI summary
An image-registration method, medium, and apparatus obtaining first and second images, generating first and second image pyramids based on the first and second images, respectively, by performing sub-sampling which reduces the length and width of each of the first and second images by half, and determining one of five directions as an optimal movement direction for a current level of the first and second image pyramids based on two images belonging to a corresponding level, updating a motion vector for the current level based on the optimal movement direction for the current level and updating a first image belonging to a level directly below the current level based on the updated motion vector for the current level, wherein the updating comprises updating a motion vector for each of a plurality of levels of the first and second image pyramids in an order from an uppermost level to a lowermost level.