Rotation-Invariant Feature Vector Generation via Frequency Domain
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image-based document retrieval methods face challenges with feature descriptors being non-robust to errors in feature point detection and rotation, and suffer from high storage requirements due to redundancy, especially when dealing with distorted query images.
Innovation Solution
A method for generating a feature vector that determines values from a set of points in an image, creates a periodic sequence based on the order of these points, and generates a feature vector from a frequency domain representation, which is invariant to rotation and reduces redundancy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional feature descriptors (e.g., LLAH) are used for document retrieval, then discrimination capability is improved, but storage requirements increase due to redundancy
Solution Approach 1:
The patent extracts only the essential information from feature descriptors by transforming them into frequency domain representations. This extraction process removes redundant components while preserving the discriminative features needed for document retrieval, thereby reducing storage requirements without sacrificing discrimination capability.
Solution Approach 2:
The patent changes the parameter representation from spatial domain to frequency domain. By applying Fourier transform to the periodic sequences derived from feature points, the feature descriptors are reparameterized in a way that eliminates redundancy while maintaining discriminative power, thus resolving the storage vs. discrimination contradiction.
2Reliability
If feature descriptors are made robust to rotation by using multiple starting points, then reliability is improved, but device complexity increases
Solution Approach 1:
The patent exploits the periodic nature of the feature sequences by transforming them into the frequency domain. The periodicity allows the system to achieve rotational robustness through frequency analysis rather than by computing multiple descriptors from different starting points, thereby maintaining reliability while reducing computational complexity.
Solution Approach 2:
The patent replaces the mechanical approach of computing multiple feature descriptors from different starting points with a mathematical transformation approach using Fourier transform. This substitution eliminates the need for iterative computation across multiple starting points while achieving the same rotational robustness, thus reducing device complexity.
3Adaptability or versatility
If local features are used to handle distorted images, then adaptability is improved, but measurement precision deteriorates due to loss of global context
Solution Approach 1:
The patent merges local feature information into a global frequency domain representation. By combining the periodic sequences from multiple local feature points and transforming them collectively, the system preserves both the local adaptability needed for distorted images and the global context necessary for accurate retrieval, thus resolving the contradiction between adaptability and measurement precision.
Data Source
AI summary
A method of generating a feature vector for an image is disclosed. Values are determined from a plurality of points in a region of the image, each of the values being determined using at least two of the plurality of points. A periodic sequence of the determined values is determined based on an order of the plurality of points. The periodic sequence is phase variant to a starting point of the ordered plurality of points, the order of the plurality of points being determined according to a predetermined rule. The feature vector for one of the points is generated from a frequency domain representation of the periodic sequence, the feature vector being invariant to rotation with respect to the plurality of points.