Per-Sample Multiple Kernel Approach for Visual Concept Learning

EURASIP Journal on Image and Video Processing. 2010;2010 DOI 10.1155/2010/461450

 

Journal Homepage

Journal Title: EURASIP Journal on Image and Video Processing

ISSN: 1687-5176 (Print); 1687-5281 (Online)

Publisher: Springer

Society/Institution: European Association for Signal Processing (EURASIP)

LCC Subject Category: Technology: Engineering (General). Civil engineering (General): Applied optics. Photonics | Technology: Electrical engineering. Electronics. Nuclear engineering: Electronics

Country of publisher: United Kingdom

Language of fulltext: English

Full-text formats available: PDF, HTML

 

AUTHORS

Ling-Yu Duan
Wen Gao
Yonghong Tian
Yuanning Li
Jingjing Yang

EDITORIAL INFORMATION

Blind peer review

Editorial Board

Instructions for authors

Time From Submission to Publication: 13 weeks

 

Abstract | Full Text

Learning visual concepts from images is an important yet challenging problem in computer vision and multimedia research areas. Multiple kernel learning (MKL) methods have shown great advantages in visual concept learning. As a visual concept often exhibits great appearance variance, a canonical MKL approach may not generate satisfactory results when a uniform kernel combination is applied over the input space. In this paper, we propose a per-sample multiple kernel learning (PS-MKL) approach to take into account intraclass diversity for improving discrimination. PS-MKL determines sample-wise kernel weights according to kernel functions and training samples. Kernel weights as well as kernel-based classifiers are jointly learned. For efficient learning, PS-MKL employs a sample selection strategy. Extensive experiments are carried out over three benchmarking datasets of different characteristics including Caltech101, WikipediaMM, and Pascal VOC'07. PS-MKL has achieved encouraging performance, comparable to the state of the art, which has outperformed a canonical MKL.