Local Feature Amount Calculation for Mobile Vision
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing local feature amount calculation methods, such as SIFT and SURF, are slow due to redundant differential value calculations and extensive floating-point operations, particularly when operated on low-specification hardware like mobile terminals.
Innovation Solution
A local feature amount calculating device that extracts feature points and near-field regions, calculates direction strengths, detects main axis directions, and acquires local feature amounts by rotating vote cell patterns instead of texture patterns, reducing the need for linear interpolation and floating-point calculations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional local feature amount calculation methods (SIFT, SURF) are used, then rotation invariance and scale invariance are achieved, but calculation speed becomes slow due to redundant differential value calculations and extensive floating-point operations
Solution Approach 1:
The patent extracts only the essential information needed for rotation invariance (main axis direction and gradient directions relative to main axis) while discarding redundant calculations. Instead of performing full differential value calculations and rotating texture patterns, the method extracts gradient directions relative to the main axis, achieving rotation invariance with significantly reduced computational complexity.
Solution Approach 2:
The patent inverts the conventional approach by not rotating the texture pattern and then calculating features, but instead calculating gradient directions relative to the main axis directly. This inversion eliminates the need for pattern rotation and subsequent differential value recalculations, maintaining rotation invariance while improving calculation speed.
2Measurement precision
If texture pattern rotation is performed to achieve rotation invariance, then accurate local feature amounts are obtained, but extensive floating-point operations and linear interpolation are required
Solution Approach 1:
The patent extracts gradient directions relative to the main axis directly from the original image coordinates without performing texture pattern rotation. This extraction approach maintains the accuracy of local feature amounts by preserving the relative angular relationships while eliminating the need for floating-point operations associated with pattern rotation and linear interpolation.
Solution Approach 2:
The patent replaces the mechanical rotation of texture patterns (which requires floating-point operations and linear interpolation) with a mathematical transformation of gradient directions. By calculating angles relative to the main axis using simple trigonometric relationships, the method achieves the same rotational alignment effect without the computational overhead of pattern rotation.
3Measurement precision
If differential value calculations are performed for each pixel in the near-field region, then accurate gradient information is obtained, but calculation time increases significantly
Solution Approach 1:
The patent performs preliminary calculation of the main axis direction from the image data before calculating gradient directions. This preliminary action allows subsequent gradient calculations to be performed relative to the established main axis, avoiding redundant differential value calculations and reducing overall computation time while maintaining gradient information accuracy.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A local feature amount calculating device includes a feature point near-field region extracting unit that extracts a feature point and a near-field region that is set near the feature point from an image, a direction strength calculating unit that calculates a direction of each pixel located within the near-field region and a strength of the direction, and a main axis direction detecting unit that detects a main axis direction of the near-field region, and a local feature amount calculating unit that acquires a local feature amount of the feature point for each pixel by giving a vote of the strength of the direction of a pixel, which is calculated by the direction strength calculating unit, to a vote cell to which a pixel of interest belongs out of vote cells of a rotated vote cell pattern that is acquired by rotating a reference vote cell pattern by an angle corresponding to the main axis direction along a corrected edge gradient direction that is acquired by correcting the edge gradient direction of the pixel that is calculated by the direction strength calculating unit in accordance with the main axis direction.