Sparse Binary Vector Encoding for Media Similarity Search
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
High-dimensional vector space data in media libraries is prone to noise, making manual and automated classification and similarity searches resource-intensive and ineffective, especially when changes in a few vector features are desired to be detected.
Innovation Solution
Encoding feature vectors into sparse binary vectors by permuting, truncating, and one-hot encoding them, allowing for efficient similarity searches and noise resilience through dot product computations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If high-dimensional feature vectors are used to represent digital media, then the representation capacity and detail information are improved, but the determination of distance between vectors is increasingly affected by compounded noise, reducing measurement precision
Solution Approach 1:
The patent extracts only the most significant features by applying discrete orthogonal transforms (DCT, DFT, wavelet transforms) to reduce high-dimensional feature vectors to lower-dimensional representations. This extraction process removes noise components while preserving essential information, resolving the contradiction between maintaining representation capacity and reducing noise impact on distance measurements.
Solution Approach 2:
The patent changes the parameter dimensionality by transforming high-dimensional feature vectors into lower-dimensional transformed feature vectors through orthogonal transforms. By adjusting the dimensionality reduction level, the system optimizes the balance between preserving meaningful information and filtering out noise that compounds in high-dimensional spaces.
2Loss of information
If the number of dimensions in a vector increases, then the representation detail is improved, but the effect of changes in a single vector component on distance between vectors becomes smaller, reducing sensitivity to feature changes
Solution Approach 1:
The patent extracts the most significant components from high-dimensional vectors through orthogonal transforms, creating a condensed representation that maintains sensitivity to meaningful changes while filtering out dimensions where noise dominates and individual component changes have negligible impact.
Solution Approach 2:
The patent transforms the feature space by applying orthogonal transforms that rotate and reorient the data in a new dimensional configuration. This dimensionality change preserves the essential variance and relationships in the data while reducing the total number of dimensions, thereby maintaining sensitivity to important feature changes.
3Measurement precision
If manual classification and indexing of digital media is performed, then classification accuracy can be maintained, but the operation becomes prohibitively expensive and time-consuming for large media collections
Solution Approach 1:
The patent replaces manual classification operations with automated computational systems that apply discrete orthogonal transforms to feature vectors. This substitution enables rapid processing of large media collections while maintaining consistent and objective classification accuracy, eliminating the time and cost constraints of manual methods.
Solution Approach 2:
The patent changes the classification approach by transforming raw feature vectors into transformed feature vectors through orthogonal transforms. This parameter transformation enables efficient automated comparison and classification operations that can process large volumes of media quickly while maintaining accuracy through mathematically rigorous distance measurements in the transformed space.
4Measurement precision
If similarity search within a large media library is performed by analyzing all entries, then search completeness is improved, but the operation requires processing resource-intensive capabilities
Solution Approach 1:
The patent extracts essential features through orthogonal transforms, creating compact representations that enable efficient comparison. This extraction allows the system to perform comprehensive similarity searches across large libraries using reduced-dimensional vectors, significantly lowering computational and energy requirements while maintaining search completeness through preserved essential information.
Solution Approach 2:
The patent changes the search implementation by transforming feature vectors into a more efficient representation space using discrete orthogonal transforms. This parameter change enables resource-efficient processing while maintaining the ability to perform complete similarity searches, as the transformed vectors preserve the necessary information for accurate distance measurements and similarity determination.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A feature vector is encoded into a sparse binary vector. The feature vector is retrieved, for example from storage or a feature vector generator. The feature vector represents a media object or other data object. One or more permutations are generated, the dimensionality of the generated permutations equivalent to the dimensionality of the feature vector. The permutations may be generated randomly or formulaically. The feature vector is permuted with the one or more permutations, creating one or more permuted feature vectors. The permuted feature vectors are truncated according to a selected window size. The indexes representing the maximum values of the permuted feature vectors are identified and encoded using one -hot encoding, producing one or more sparse binary vectors. The sparse binary vectors may be concatenated into a single sparse binary vector and stored. The sparse binary vector may be used in the similarity search, indexing or categorization of media objects.