Circular Distribution Local Image Feature Descriptors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing local image feature descriptors, such as SIFT, are less effective in accurately matching images that undergo geometric and photometric transformations like blurring, zoom/rotation, lighting changes, and JPEG compression due to their limited representation of oriented gradients.
Innovation Solution
The enhancement of histogram-based local image feature descriptors by incorporating circular distribution information modeled through a mixture of circular normal distributions learned via the expectation maximization function, which provides a set of circular means and variances to accurately represent the distribution of oriented gradients.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional histogram-based local image feature descriptors are used, then the computational complexity is low and the method is simple, but the matching accuracy deteriorates under geometric and photometric transformations
Solution Approach 1:
The patent segments the histogram representation by introducing circular distribution information that divides the gradient orientation space into meaningful segments. The circular normal distributions segment the gradient data into clustered groups with distinct means and variances, allowing the descriptor to capture fine-grained distribution patterns while maintaining computational efficiency through a structured approach.
Solution Approach 2:
The patent adds a new dimension to the traditional histogram by incorporating circular distribution parameters (means and variances from circular normal distributions). This transforms the flat histogram into a multi-dimensional representation that includes not just gradient counts but also distribution characteristics, enhancing the descriptor's ability to distinguish transformed images without excessive complexity.
2Measurement precision
If the representational power of local image feature descriptors is increased, then the matching accuracy under transformations improves, but the computational cost increases
Solution Approach 1:
The patent applies partial action by selectively enhancing the histogram with circular distribution information only where needed - specifically in the orientation gradient domain. Rather than completely transforming the descriptor framework, it adds circular statistical moments (means and variances) to existing histogram bins, providing enhanced representational power for transformation robustness while avoiding the computational burden of complete descriptor redesign.
Solution Approach 2:
The patent changes parameters by introducing circular statistical parameters (means and variances from circular normal distributions) to characterize gradient orientations. These parameter additions enrich the descriptor's ability to represent orientation distributions under transformation without requiring fundamental changes to the computational pipeline, maintaining efficiency while improving accuracy.
Data Source
AI summary
A method and system characterizes an image of an object. A plurality of interest points are detected within a first image and a local image feature descriptor is built for at least some of the interest points, including mapping information about the interest points according to at least circular distribution information.