摘要
针对通用型图像检索面对不同图像数据难以事先确定合适图像特征的问题,提出一种基于互信息(MI)作为相关性测度的相似度选择的排序支持向量机精化最终结果的检索方法。该方法采取融合大量全局特征以及局部特征的策略,通过在少量的训练集上进行线下相似度选择,在提供良好检索准确性的基础上大幅减少后续检索中使用的特征数量,并且通过排序支持向量机进一步提高检索准确性。在自然图像Wang数据集以及医学图像IRMA数据集上进行实验,结果表明该方法优于常用的单一最好相似度以及典型的"和规则"组合方法,分别比准确性第二好的方法提高了13.3%和96.7%。公共数据集的实验结果表明,所提方法能够提高通用型图像检索系统的准确性。
For general-purpose image retrieval, it is difficult to determine the appropriate image features for an unknown dataset in advance. Therefore, a new retrieval method based on Mutual Information（ MI） and rank-SVM（ Support Vector Machine） was proposed, combining similarities via mutual information selection and rank-SVM refining steps. Most relevant similarities were selected by using a small amount of training set and computing off-line. Rank-SVM was used to further refine the initial retrieval result. Through experiments on different public image sets Wang and IRMA, this method is proved to be superior to conventional single best feature＇s similarity as well as the typical ＂ sum rule＂. The accuracy increments than the second best method are 13. 3% on Wang data set and 96. 7% on IRMA data set respectively. The results on public data sets show that the proposed method can improve the accuracy of general-purpose image retrieval system.
出处
《计算机应用》
CSCD
北大核心
2017年第A01期165-168,172,共5页
journal of Computer Applications
基金
国家民委科研项目(14BFZ007)
国家自然科学基金资助项目(61561002)
宁夏区自然科学基金资助项目(NZ15105)
关键词
基于内容的图像检索
相似度组合
最大依赖
互信息
排序支持向量机
Content-Based Image Retrieval（CBIR）
similarity combination
max-dependency
Mutual Information（MI）
rank-SVM（Support Vector Machine）
