Related Experiment Videos
Asymmetric bagging and random subspace for support vector machines-based relevance feedback in image retrieval
Dacheng Tao1, Xiaoou Tang, Xuelong Li
1School of Computer Science and Information Systems, Birkbeck, University of London, Malet Street, London WC1E 7HX, UK. dacheng@dcs.bbk.ac.uk
Summary
Support Vector Machines (SVM) in content-based image retrieval (CBIR) struggle with limited positive feedback. This study introduces asymmetric bagging and random subspace SVM (ABRS-SVM) to enhance performance by addressing SVM instability and overfitting.
Area of Science:
- Computer Science
- Machine Learning
- Information Retrieval
Background:
- Support Vector Machines (SVM) are prevalent in content-based image retrieval (CBIR) relevance feedback.
- SVM performance degrades with few positive feedback samples due to instability, hyperplane bias, and overfitting.
Purpose of the Study:
- To enhance the performance of SVM-based relevance feedback in CBIR.
- To overcome the limitations of SVM with small positive feedback datasets.
Main Methods:
- Proposed an asymmetric bagging-based SVM (AB-SVM) to address SVM instability and hyperplane bias.
- Introduced a random subspace SVM (RS-SVM) to combat overfitting in high-dimensional feature spaces.
- Integrated AB-SVM and RS-SVM into an asymmetric bagging and random subspace SVM (ABRS-SVM) for comprehensive improvement.
Main Results:
- The developed ABRS-SVM mechanism effectively addresses the key challenges in SVM-based relevance feedback.
- The integrated approach demonstrates improved relevance feedback performance in CBIR.
Conclusions:
- ABRS-SVM offers a robust solution for improving CBIR relevance feedback accuracy, particularly when dealing with limited labeled data.
- This research contributes a novel ensemble method for enhancing SVM-based retrieval systems.