Image Stitching Using Wavelet Featureness and Correlation Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing digital image stitching techniques require significant memory and CPU resources, especially when dealing with large numbers of high-resolution images, leading to memory overload and slow stitching processes, while maintaining good quality and speed remains a challenge.
Innovation Solution
A method that calculates featureness and correlation coefficients to determine optimal stitching paths with minimal memory and CPU usage, employing wavelet transforms and correlation analysis to efficiently align and combine images, using a reliability-oriented global max-span tree for determining the best connection route.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional stitching techniques load all images into memory for stitching, then stitching quality can be maintained, but memory consumption becomes enormous when dealing with large numbers of high-resolution images
Solution Approach 1:
The patent divides the stitching process into multiple passes, processing a limited number of images (e.g., 5-10 images) in each pass rather than loading all images simultaneously. This segmentation allows high-resolution image stitching while keeping memory consumption within acceptable limits by processing images in manageable batches across multiple sequential passes.
2Quantity of substance
If images are saved as digital files and reloaded for stitching to avoid large memory usage, then memory consumption is reduced, but the repeated image loading slows down the whole stitching process
Solution Approach 1:
The patent performs preliminary processing by down-sampling images to lower resolution versions first, using these reduced images to calculate feature points and determine stitching parameters in advance. This preliminary action allows the actual high-resolution stitching to proceed more efficiently without repeated loading, as the alignment parameters are pre-determined from the lower-resolution previews.
3Productivity
If feature points are calculated from lower resolution versions and then refined recursively, then computation speed improves, but the recursive search process increases CPU usage
Solution Approach 1:
The patent applies partial refinement by performing feature matching first on down-sampled lower-resolution images to obtain initial alignment parameters, then applying only necessary refinements to these parameters using the original high-resolution images. This partial action approach achieves good stitching results without performing exhaustive recursive searches on all high-resolution data, thereby reducing CPU usage while maintaining acceptable computation speed.
Data Source
AI summary
The presently claimed invention provides a method for stitching a plurality of images together in a way with least memory and CPU usage, and minimum file input and output, while still possessing fast computation speed to avoid post-scan delay for whole slide viewing and good stitching quality. The method of the present invention comprises the steps of calculating featureness of each candidate strip of a image by applying a mathematical transformation, calculating offset of the strip with correlation maximum location, calculating stitching reliability of the candidate strip from the pre-defined weighted function of its featureness and the correlation coefficients of each matching block, and determining optimal stitching path with stitching reliability.