ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2016, Vol. 53 ›› Issue (5): 1043-1052.doi: 10.7544/issn1000-1239.2016.20150070

• 人工智能 • 上一篇    下一篇

基于词典优化与空间一致性度量的目标检索

赵永威1,周苑2,李弼程3   

  1. 1(武警工程大学电子技术系 西安 710000); 2(河南工程学院计算机学院 郑州 451191); 3(解放军信息工程大学信息系统工程学院 郑州 450002) (zhaoyongwei369@163.com)
  • 出版日期: 2016-05-01
  • 基金资助: 
    国家自然科学基金项目(60872142,61301232)

Object Retrieval Based on Enhanced Dictionary and Spatially-Constrained Similarity Measurement

Zhao Yongwei1, Zhou Yuan2, Li Bicheng3   

  1. 1(Department of Electronic Technology, CAPF Engineering University, Xi’an 710000); 2(School of Computer Science, Henan University of Engineering, Zhengzhou 451191); 3(Institute of Information System Engineering, PLA Information Engineering University, Zhengzhou 450002)
  • Online: 2016-05-01

摘要: 基于视觉词典模型(bag of visual words model, BoVWM)的目标检索存在时间效率低、词典区分性不强的问题,以及由于空间信息的缺失及量化误差等导致的视觉语义分辨力不强的问题.针对这些问题,提出了基于词典优化与空间一致性度量的目标检索方法.首先,该方法引入E\+2LSH(exact Euclidean locality sensitive hashing)过滤图像中的噪声和相似关键点,提高词典生成效率和质量;然后,引入卡方模型(chi-square model, CSM)移除词典中的视觉停用词增强视觉词典的区分性;最后,采用空间一致性度量准则进行目标检索并对初始结果进行K-近邻(K-nearest neighbors, K-NN)重排序.实验结果表明:新方法在一定程度上改善了视觉词典的质量,增强了视觉语义分辨能力,进而有效地提高目标检索性能.

关键词: 目标检索, 视觉词典模型, 精确欧氏位置敏感哈希, 空间一致性度量, 卡方模型

Abstract: Bag of visual words model based object retrieval methods have several problems, such as low time efficiency, the low distinction of visual words and the weakly visual semantic resolution because of missing spatial information and quantization error. In this article, an object retrieval method based on enhanced dictionary and spatially-constrained similarity measurement is proposed aiming at the above problems. Firstly, E\+2LSH (exact Euclidean locality sensitive hashing) is used to identify and eliminate the noise key points and similar key points, consequently, the efficiency and quality of visual words are improved; Then, the stop words of dictionary are eliminated by chi-square model (CSM) to improve the distinguish ability of visual dictionary; Finally, the spatially-constrained similarity measurement is introduced to accomplish object retrieval, furthermore, a robust re-ranking method with the K-nearest neighbors of the query for automatically refining the initial search results is introduced. Experimental results indicate that the quality of visual dictionary is enhanced, and the distinguish ability of visual semantic expression is effectively improved and the object retrieval performance is substantially boosted compared with the traditional methods.

Key words: object retrieval, bag of visual words model, exact Euclidean locality sensitive hashing(E\+2LSH), spatially-constrained similarity measure, chi-square model (CSM)

中图分类号: