Depth Normalization for Feature Point Search Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image search methods using local feature amounts struggle with high precision when images are subjected to out-of-plane rotations beyond 30 degrees, due to changes in pixel ranges and depth values, leading to degraded search accuracy.
Innovation Solution
An image processing apparatus and method that extracts feature points from images with changing depth values, normalizes depth values using a transformation matrix to generate images for feature amount calculation, and calculates feature amounts invariant to rotations, thereby reducing the impact of out-of-plane rotations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If local feature amounts are calculated using conventional pixel ranges, then the method is simple and fast, but search accuracy degrades for images rotated beyond 30 degrees
Solution Approach 1:
The patent applies parameter changes by transforming pixel coordinates through a normalization transformation matrix that accounts for depth variations. This transforms the pixel range selection from a fixed conventional approach to a dynamic approach that adapts to the specific depth profile of each local region, thereby maintaining search accuracy for rotated images while managing processing complexity through efficient matrix operations
Solution Approach 2:
The patent introduces a depth dimension to the traditional 2D pixel range calculation. By considering the depth profile of pixels in the vertical direction and incorporating this third dimension into the normalization transformation, the system can accurately identify corresponding points even when images are rotated out-of-plane, significantly improving search accuracy beyond the conventional 30-degree limitation
2Measurement precision
If depth values are normalized using transformation matrices, then search accuracy for rotated images improves, but processing time increases
Solution Approach 1:
The patent applies preliminary action by pre-calculating and storing depth profile information for local regions before the actual feature matching process. The normalization transformation matrices are computed in advance based on the depth profiles, so that during the search process, only simple matrix multiplications are needed rather than performing complex depth analysis in real-time, thus reducing processing time while maintaining accuracy
Solution Approach 2:
The patent optimizes the normalization transformation by changing the approach from computing full 3D transformations to using simplified 2D normalization matrices that operate on pixel coordinates directly. This parameter change reduces the computational burden during image search while preserving the accuracy benefits of depth-aware normalization
3Reliability
If conventional pixel ranges are used without depth normalization, then processing is fast and simple, but feature amounts become sensitive to out-of-plane rotations
Solution Approach 1:
The patent changes the parameters used for defining pixel ranges from fixed conventional ranges to depth-normalized ranges that adapt to the specific depth profile of each local region. This transformation makes the feature amount calculation robust to out-of-plane rotations by normalizing the depth values, thereby improving reliability without requiring overly complex transformations
Solution Approach 2:
The patent creates a normalized version of the image region by applying the depth normalization transformation to generate a transformed image. This transformed image serves as a copy that preserves the essential features while removing depth-related variations, allowing for reliable feature matching across rotated images without modifying the original image data
Data Source
AI summary
A feature point is extracted from an input image including an image region for which depth values of pixels change consecutively. A transformation that normalizes depth values of pixels of a vicinity of the feature point with respect to a region of at least a portion of the input image is set as a normalization transformation, and an image for a feature amount calculation is generated by performing the normalization transformation on a pixel position of the feature point. A feature amount is calculated from the image for the feature amount calculation.