Making QBIC Faster
(
printing version
)
Making QBIC Faster
Overview
The QBIC System
Content-based Image Retrieval
The Problem
High-dimensional Nearest Neighbours Search
The Problem (Re-stated)
The Context
The Costs
Cost Parameters
Analysing Costs
Measuring Costs
Current Approach
Specific Goals for a Solution
Some Potential Solutions
Tree-based Methods
Insertion with Trees
Query with Trees
Why Trees Don't Work (for
d > 10
)
Signature-based Methods
(VA-Files)
VA-File Signatures
Insertion with VA-Files
Query with VA-Files
Why VA-Files Don't Work (for QBIC)
Projection-based Methods
(C-curves)
Proposed Approach: CurveIx
Curve Mapping Functions
Data Structures for CurveIx
Database Construction
Example: Search with 5 Curves
Finding
k
-NN
(Simple Approach)
CurveIx vs. Linear Scan
Implementation
Example for
build
Program
Example for
query
Program
C
i
Values as Keys
Query Algorithm
Performance Analysis
Experiments
Sample Comparison
Experimental Results
Results: Size vs. Accuracy
Results: Costs
Other Optimisations for QBIC Indexing
Caching
Compression
Clustering
Conclusions
References/Further Reading
Produced: 28 Aug 98