电子学报 ›› 2010, Vol. 38 ›› Issue (9): 2101-2106.

• 学术论文 • 上一篇    下一篇

基于混合学习框架的SVM反馈算法研究

邬俊, 鲁明羽, 刘闯   

  1. 大连海事大学信息科学技术学院,辽宁大连 116026
  • 收稿日期:2009-07-16 修回日期:2010-03-09 出版日期:2010-09-25 发布日期:2010-09-25

SVM-Feedback Scheme Within Hybrid Learning Framework for Image Retrieval

WU Jun, LU Ming-yu, LIU Chuan   

  1. School of Information Science & Technology, Dalian Maritime University, Dalian, Liaoning 116026, China
  • Received:2009-07-16 Revised:2010-03-09 Online:2010-09-25 Published:2010-09-25

摘要: 基于支持向量机(SupportVectorMachine,SVM)理论的相关反馈技术是可有效提高图像检索性能的重要手段之一。然而,大多数SVM反馈算法普遍受到小样本问题的制约。本文综合了集成学习、半监督学习和主动学习三种方法的技术特点,提出一种混合学习框架下的SVM反馈算法。该算法在Boosting迭代过程中使用了未标记图像,以增加个体SVM之间的差异,从而获得高效的集成学习模型。同时,高效的集成学习模型更有利于寻找富有信息(mostinformative)图像,从而也提高了用户主动反馈的效率。实验结果及对比分析表明,混合学习策略可有效改进相关反馈的性能。

关键词: 图像检索, 相关反馈, 支持向量机, 混合学习

Abstract: Relevance feedback plays an important role for enhancing content-based image retrieval (CBIR).Among various methods,support vector machine (SVM) based relevance feedback technique has drawn substantial research attention.However,most SVM-feedback approaches are challenged by the small example issue.This paper presents a SVM-feedback scheme within the hybrid learning framework that integrates the merits of ensemble learning,semi-supervised learning and active learning in order to achieve strong generalization.Concretely,in each round of boosting iterations,unlabeled images are exploited to augment the diversity among component SVM learners,and thus a powerful ensemble learning model is constructed.Conversely,the enhanced ensemble learning model is helpful to identify the most informative images which are used for active feedback.Experimental results,including comparison analysis,show that the hybrid learning framework is efficient to improve relevance feedback performance.

Key words: image retrieval, relevance feedback, support vector machine, hybrid learning

中图分类号: