Video Detection Using Disjoint Feature Space Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing content-based retrieval systems face challenges in efficiently and accurately monitoring and managing digital video content due to high-dimensional feature spaces, which lead to intractable K-NN queries and loss of accuracy in feature space transformations.
Innovation Solution
A content-based retrieval system that uses nested, disjoint feature spaces for digital data representations, employing sequential feature data set extraction, spatial and temporal feature analysis, and error metric-based matching to efficiently compare and classify digital video content, enabling scalable and accurate video detection and monitoring.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If hierarchical tree structures (R-tree, R+-tree, R*-tree) are used to index K-NN queries in high-dimensional feature spaces, then query organization is improved, but query tractability deteriorates when dimension K exceeds approximately 20
Solution Approach 1:
The patent segments the high-dimensional feature space into multiple lower-dimensional subspaces using disjoint hyperspheres. Each subspace is indexed independently, avoiding the curse of dimensionality that plagues traditional hierarchical structures when K>20. This segmentation allows K-NN queries to be tractable by limiting search to relevant subspaces rather than the entire high-dimensional space.
Solution Approach 2:
The patent transforms the problem from searching in a single high-dimensional space to searching across multiple lower-dimensional subspaces organized in a hierarchical structure. By distributing features across different dimensional subspaces and using hyperspherical boundaries, the system achieves efficient indexing without the exponential complexity growth that occurs in traditional high-dimensional tree structures.
2Productivity
If feature space dimension is significantly reduced using fast approximation of Karhunen-Loeve Transform, then processing speed is improved, but K-NN query accuracy deteriorates
Solution Approach 1:
Instead of reducing the entire feature space to fewer dimensions, the patent segments the feature space into multiple subspaces that collectively preserve the original dimensionality. Each subspace maintains sufficient dimensional information for accurate K-NN queries while being small enough for efficient processing. This avoids the accuracy loss inherent in aggressive dimensionality reduction.
Solution Approach 2:
The patent applies different processing characteristics to different regions of the feature space through the disjoint hypersphere subspaces. Each subspace can be optimized independently for its local characteristics, maintaining high query accuracy locally while achieving overall system efficiency through parallel processing of multiple subspaces.
3Device complexity
If traditional tree structures partition feature space into irregular convex subspaces, then spatial organization is improved, but K-NN query efficiency deteriorates in high-dimensional spaces
Solution Approach 1:
The patent replaces irregular convex subspaces with disjoint hyperspherical subspaces. The hyperspherical geometry provides more uniform coverage of the feature space and creates well-defined boundaries that facilitate efficient query processing. The spherical shape allows for constant-distance searches within each subspace, improving K-NN query efficiency compared to irregular convex partitions in high-dimensional spaces.
Data Source
AI summary
A video detection system and method compares a queried video segment to one or more stored video samples. Each of the queried video segments and stored video samples can be represented by respective digital image sets. A first and second comparison comprises comparing a set of low and high resolution temporal and spatial statistical moments in a COLOR9 space, and eliminating file digital image sets that do not match the queried digital image set. A third comparison generates a set of matching files by comparing a set of wavelet transform coefficients in a COLOR9 space. RGB bit-wise registration and comparison of one or more subframes of specific frames in the queried digital image set to a corresponding set of matching file subframes determines queried subframe changes. In the event of a change in a queried subframe, the changed subframe is added to the set of matching file subframes.