Image Hash Code Determination via Delaunay Region Canonical Form
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing content-based image retrieval methods face challenges in efficiently determining the distance to a plane in high-dimensional spaces, leading to slow registration processes and false-negative matches due to the computational complexity of SVD methods.
Innovation Solution
A method that determines a hash code for an image by transforming a Delaunay region to a canonical form, setting specific coordinates of the normal to non-zero values, and using these to calculate the hash code based on the distance, thereby improving the efficiency of image retrieval in high-dimensional spaces.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If SVD method is used to calculate the normal of a plane in high-dimensional space, then measurement precision is maintained, but device complexity and computation time increase significantly
Solution Approach 1:
The patent segments the normal calculation process into two distinct phases: (1) an offline preprocessing phase where the projection direction is pre-computed using SVD for high precision, and (2) an online query phase where only simple dot product operations are performed. This segmentation allows the complex SVD computation to be done once during system setup, while fast operations are used during actual image retrieval, thus resolving the contradiction between precision and computational complexity.
Solution Approach 2:
The patent applies preliminary action by pre-computing the projection direction vector through SVD decomposition of the covariance matrix during an offline training phase. This pre-computed projection direction is then stored and reused for all subsequent image retrievals, eliminating the need to perform expensive SVD calculations during online queries. The preliminary computation ensures measurement precision is maintained while avoiding repeated complex calculations.
2Measurement precision
If exhaustive search method is used for image retrieval, then measurement precision is maintained, but productivity decreases for large databases
Solution Approach 1:
The patent substitutes the mechanical exhaustive search approach with a mathematical projection-based method. Instead of computing Euclidean distances between query and all database images (mechanical comparison), the system projects both onto a pre-computed direction vector and uses simple dot product operations. This substitution maintains similarity search accuracy while dramatically improving retrieval speed for large databases by replacing O(N) distance computations with O(1) projection operations.
Solution Approach 2:
The patent changes the parameter space by transforming the image data from the original high-dimensional space into a projected 1D space along a pre-computed direction vector. This parameter transformation allows the system to maintain the essential similarity relationships while enabling faster comparison operations. The projection parameter (direction vector) is carefully chosen through offline SVD to preserve the most discriminative information, thus maintaining measurement precision while improving productivity.
3Reliability
If hash-based methods with multiple probes are used to overcome false-negative matches, then reliability improves, but loss of time increases due to multiple probes
Solution Approach 1:
The patent introduces a projection direction vector as an intermediary that maps high-dimensional image features onto a 1D scale. This intermediary transformation preserves the similarity structure of the data while enabling direct comparison operations. By using this intermediary projection, the system achieves reliable match detection without needing multiple probes, as the projection inherently maintains the ordering and relative distances of similar images, thus improving reliability while reducing time loss.
Data Source
AI summary
A method of determining a hash code representing a portion of an image, is disclosed. A Delaunay region (e.g., 450) enclosing an image feature point (e.g., 210) representing at least the portion of the image is determined. The Delaunay region is determined from A* lattice points. A mapping transforming the Delaunay region to a predetermined canonical form is determined A point of the Delaunay region is received. The received point defines a plane containing the A* lattice points of the Delaunay region excluding the received point. A normal of the plane is determined by setting at least two co-ordinates of the normal to predetermined non-zero values, the two co-ordinates being selected according to the determined mapping. The hash code representing a portion of the image is determined according to a distance determined using the normal.