by p. muneesawang and l. guan ieee transactions...
TRANSCRIPT
by P. Muneesawang and L. GuanIEEE Transactions on Multimedia, 2004
Joseph Lee
1. Backgroundg
2. General Framework
d l3. Previous Models
4. Proposed RBF Model
5. Learning Strategy
E perimental Res lts6. Experimental Results
7. Conclusion
Content-based Image Retrieval (CBIR)◦ use colors, shapes, textures → rely on image itself◦ without image content → rely on metadata
Content-based Image Retrieval (CBIR)◦ use colors, shapes, textures → rely on image itself◦ without image content → rely on metadata
Relevance Feedback◦ User refines the result by marking images as
relevant or non-relevant and repeats search.
Image Similarity
Image Similarity
Examples of Image Similarity Problems
MARS-1 RBF
Ideal Case
Ideal Case – User Query
Ideal Case – Similarity Measure
Real Case
Real Case – Inaccurate Query
Real Case – Inaccurate Query
We should refine Query
QueryQuery
We should refine Query & Metric
Metric
QueryQuery
Query Reformulation Model
◦ Relevance feedback → learn query representation
Adaptive Metric Model
◦ Relevance feedback → learn similarity function
Learning user perception
◦ MARS-1◦ MARS 1
◦ MARS-2
◦ OPT-RF
Assumption◦ same distance gives same degree of similarity
Learning user perception
◦ MARS-1◦ MARS 1
◦ MARS-2
◦ OPT-RF
Assumption◦ same distance gives same degree of similarity
Learning user perception
◦ MARS-1◦ MARS 1
◦ MARS-2
◦ OPT-RF
Assumption◦ same distance gives same degree of similarity
Learning user perception
◦ MARS-1◦ MARS 1
◦ MARS-2
◦ OPT-RF
Assumption◦ same distance gives same degree of similarity
Learning user perception
◦ MARS-1◦ MARS 1
◦ MARS-2
◦ OPT-RF
Assumption◦ same distance gives same degree of similarity
CBIR◦ online learning◦ two-class problem
CBIR◦ online learning◦ two-class problem
Example Vectors
Voronoi Vectors
Voronoi Cells
Learning Vector Quantization
Learning Vector Quantization
Learning Vector Quantization
Learning Vector Quantization
Modified LVQ (Model 1)
Modified LVQ (Model 1)
Modified LVQ (Model 1)
Modified LVQ (Model 1)
Modified LVQ (Model 2)
Effects of Positive and Negative Learning
◦ Relevant samples → common interest
◦ Non-relevant samples → specific interest
Effects of Positive and Negative Learning
◦ Relevant samples → common interest
◦ Non-relevant samples → specific interest
Effects of Positive and Negative Learning
◦ Relevant samples → common interest
◦ Non-relevant samples → specific interest
Effects of Positive and Negative Learning
◦ Relevant samples → common interest
◦ Non-relevant samples → specific interest
Selection of RBF Width
Selection of RBF Width
Selection of RBF Width
◦◦
Average Precision Rate (%)
Non-linear model for similarity evaluation
Learning from positive and negative samples
http://videolectures.net/minh_hoai_nguyen/http://www mqasem net/vectorquantization/vq htmlhttp://www.mqasem.net/vectorquantization/vq.html