Lattice-Based Hashing for Image Retrieval Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing content-based image searching methods face challenges in efficiently retrieving similar images from large databases with high-dimensional feature vectors, particularly due to high false-negative matches and increased memory and time requirements with existing hashing approaches.
Innovation Solution
A method and system that select a plurality of lattice points in a multidimensional lattice to form a smallest enclosing region about a feature vector, determining a lattice point based on a distribution criteria, and assigning the feature vector to that lattice point, thereby linking a hash code to the image and minimizing false-negatives by considering feature vectors within a query radius.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If LSH uses multiple hash functions to reduce false-negative matches, then accuracy is improved, but memory requirement and search time increase
Solution Approach 1:
The patent segments the high-dimensional feature space into multiple hash tables, each handling a subset of dimensions. This allows the system to achieve accurate retrieval without requiring all hash functions to be loaded into memory simultaneously, thus reducing peak memory requirements while maintaining retrieval accuracy through coordinated searching across segmented tables.
Solution Approach 2:
The patent transforms the problem from handling all dimensions simultaneously in a single hash table to processing dimensions in separate groups across multiple hash tables. This dimensional segmentation enables the system to achieve the same retrieval accuracy with reduced memory footprint by distributing the storage burden across multiple smaller tables.
2Reliability
If LSH uses multiple hash functions to reduce false-negative matches, then accuracy is improved, but search time increases
Solution Approach 1:
By segmenting the hash functions into groups assigned to different hash tables, the system can search multiple tables in parallel rather than sequentially processing a single large table. This segmentation enables faster search times while maintaining the accuracy benefits of using multiple hash functions.
Solution Approach 2:
The patent allows for partial searching of hash tables based on query requirements. Instead of always searching all hash tables completely, the system can perform partial searches that stop early when sufficient results are found, reducing average search time while maintaining accuracy through the distributed hash table structure.
3Reliability
If exhaustive search is used to ensure accurate retrieval, then accuracy is improved, but search speed deteriorates
Solution Approach 1:
The patent performs preliminary hashing of query vectors into multiple hash tables before conducting the full search. This preliminary action organizes the search space in advance, allowing the system to quickly narrow down candidate matches across multiple tables and then perform exhaustive verification only on the reduced candidate set, thereby maintaining accuracy while improving search speed.
4Productivity
If hash-based strategies are used to improve search speed, then productivity is improved, but false-negative matches increase
Solution Approach 1:
The patent merges multiple hash table search results to compensate for false negatives in individual tables. By combining the candidate sets from multiple hash tables that use different hash functions, the system ensures that matches missed by one hash function are likely to be caught by others, thereby reducing false-negative rates while maintaining the speed benefits of hash-based searching.
Data Source
AI summary
A system and method for linking a hash code to a portion of an image. A plurality of lattice points is selected in a multidimensional lattice to form a smallest enclosing region about a feature vector representing the portion of the image and a lattice point is determined from the selected plurality of lattice points according to a distribution criteria. The determined lattice point is common to the smallest enclosing region and a region of the lattice adjacent to the smallest enclosing region located within a query radius distance of the feature vector. When the feature vector is located within the query radius of a query vector the feature vector is considered a match. The method assigns the feature vector to the determined lattice point and stores a link between a hash code associated with the determined lattice point and the portion of the image.