Rotation Invariant Feature Descriptors for Digital Image Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for determining feature descriptors for points of interest in digital images fail to ensure rotation invariance, leading to higher mismatches when images are captured at different orientations or contain congruent points, due to assumptions of upright camera orientation and lack of invariance to rotation changes.
Innovation Solution
The method transforms digital images to align points of interest with a principal point, applying pan and tilt angles to create a common hypothetical image plane, and determines rotation invariant feature descriptors that are pan, tilt, and roll invariant, ensuring consistent description across different orientations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing approaches extract points of interest and form feature descriptors based on neighborhood pixels, then feature extraction and description can be performed, but rotation invariance is lost leading to higher mismatches when images are captured at different orientations
Solution Approach 1:
The patent applies parameter changes by transforming the feature descriptor from a fixed neighborhood-based representation to a rotation-invariant representation. This is achieved by computing the orientation of the point of interest relative to the image center and using this orientation information to rotate the feature descriptor or the image patch, thereby making the descriptor invariant to rotation. The key parameter changed is the orientation angle, which allows the same physical point to have consistent descriptors across different camera orientations.
2Ease of manufacture
If existing approaches assume upright camera orientation, then feature descriptor determination is simplified, but mismatches increase when images contain congruent points or are captured at different orientations
Solution Approach 1:
The patent applies preliminary action by pre-computing the orientation of each point of interest relative to the image center before extracting the feature descriptor. This orientation information is stored and used during the matching process to adjust and align descriptors from images captured at different orientations. By preparing this orientation data in advance, the system can handle rotated images without requiring complex real-time computation, thus maintaining simplicity while improving accuracy.
3Measurement precision
If feature descriptors are formed around points of interest using neighborhood pixels, then local feature information is captured, but invariance to changes in viewing direction and rotation is not achieved
Solution Approach 1:
The patent applies asymmetry by introducing an asymmetric transformation based on the orientation angle of each point of interest. Instead of using a symmetric, rotation-invariant neighborhood aggregation, the method computes the angle between the point and the image center, then uses this asymmetric angle information to rotate or transform the feature descriptor. This asymmetric approach allows the descriptor to adapt to the specific orientation of each point, achieving consistency across different viewing directions while maintaining precise local feature information.
Data Source
AI summary
A system and method for determining rotation invariant feature descriptors for points of interest in digital images for image matching are disclosed. In one embodiment, a point of interest in each of two or more digital images is identified. Further, the digital images are transformed to change location of the point of interest in each of the digital images to a principal point. Furthermore, a rotation invariant feature descriptor is determined for the point of interest in each of the transformed digital images for image matching.