Image Feature Database Indexing With Unique Identifier Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for creating and searching image databases face challenges due to high-dimensional image features, which lead to inefficient storage and retrieval, particularly in large datasets, as they require complex tree structures and extensive computational resources, making it difficult to store and quickly search for similar images.
Innovation Solution
A method that normalizes feature vectors into unique identifiers, allowing for efficient storage and retrieval by using two databases: one for storing identifiers with associated object IDs and another for unique identifiers, enabling fast search and reducing false positives through unique identifier matching.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-dimensional image features are used for image recognition, then the accuracy of object identification is improved, but the complexity of storage and retrieval operations increases significantly
Solution Approach 1:
The patent divides the image database into multiple clusters, with each cluster containing images that share similar features. This segmentation allows the system to manage high-dimensional features more efficiently by organizing them into manageable groups, reducing the complexity of storage and retrieval operations while maintaining identification accuracy.
Solution Approach 2:
The patent introduces cluster centers as intermediary representations that summarize the characteristics of each cluster. Instead of directly comparing query features with all individual image features, the system uses cluster centers as mediators to guide the search process, significantly reducing computational complexity while preserving recognition accuracy.
2Quantity of substance
If conventional database structures are used to store image features, then the storage capacity is sufficient, but the search speed decreases due to the need for complex tree traversals
Solution Approach 1:
The patent segments the large image database into multiple smaller clusters, each with its own compact representation. This segmentation enables faster searching by limiting the search scope to relevant clusters rather than traversing the entire database, thereby improving search speed while maintaining adequate storage capacity.
Solution Approach 2:
The patent performs preliminary clustering and compression of image features during the database construction phase. By pre-organizing images into clusters and computing cluster centers in advance, the system eliminates the need for complex tree traversals during query operations, significantly improving search speed while preserving storage capacity.
3Adaptability or versatility
If image features are stored in high-dimensional vectors, then the representational power is improved, but the memory requirements increase and false positives occur
Solution Approach 1:
The patent extracts the essential characteristics from high-dimensional image feature vectors by computing cluster centers that represent each cluster. This extraction process removes redundant information while preserving the core representational power, thereby reducing memory requirements while maintaining adaptability for various image recognition tasks.
Solution Approach 2:
The patent transforms high-dimensional feature vectors into lower-dimensional cluster center representations by changing the parameter space. This parameter transformation reduces memory requirements by compressing the data while preserving the essential information needed for accurate image recognition and reducing false positives.
4Productivity
If clustering is performed to group similar images, then the search efficiency is improved, but the computational time for creating and updating the database increases
Solution Approach 1:
The patent performs clustering as a preliminary action during the database construction phase, organizing images into clusters and computing cluster centers before actual search operations begin. This preliminary clustering improves search efficiency during query operations, and the computational time is amortized over the lifetime of the database, making the trade-off worthwhile.
Data Source
Figure 1
Figure 2
AI summary
The method involves providing normalized feature vector and associated object identity (ID) for object e.g. image captured by camera (1). The identifier is extracted from several feature vectors, and is stored with associated object ID for each object in database. The feature vector for each object is selectively stored in database. Independent claims are included for the following: (1) device for creating object database; (2) method for obtaining object from object database; and (3) device for obtaining object from object database.