基于ISODATA聚类的词汇树图像检索算法  

Image Retrieval of Vocabulary Tree Method Based on ISODATA

作　　者：张婷[1] 戴芳[1] 郭文艳[1] 

机构地区：[1]西安理工大学理学院,西安710054

出　　处：《计算机科学》2014年第B11期123-127,共5页Computer Science

基　　金：国家重大研究计划培育项目(91120014);陕西省教育厅科研计划项目(12JK0534)资助

摘　　要：词汇树图像检索是一种基于视觉关键词结构的高效的图像检索算法。该算法在特征提取和聚类过程中分别采用SIFT算法和K-means算法。然而,K-means算法对初值比较依赖,当聚类个数未知时,聚类易出现强分现象,且SIFT算法易造成数据溢出和增加检索时间。对此,给出了两种新的特征提取方法,分别称为SIFT_CRONE特征和Color_HU特征,同时引入了ISODATA算法对特征进行聚类。SIFT_CRONE特征提取方法基于SIFT算法确定图像的关键点,采用CRONE算子计算关键点周围像素的梯度,对关键点进行向量描述,其优点是既保持了SIFT特征的优点又减少了检索时间。Color_HU特征是利用SIFT确定关键点和有效区域,对关键点的邻域提取该感兴趣区域的颜色直方图和HU矩特征,降低特征维数,缩短检索时间。在使用ISODATA算法时,设计了一种自适应参数确定算法。实验结果表明,ISODATA算法克服了K-means对初值的依赖,当聚类个数未知时有较好的聚类效果;两种新特征有各自的特点,均可以缩短图像的检索时间,提高检索效率。Vocabulary tree image retrieval is a kind of efficient image retrieval algorithm based on the structure of visual words.It employes SIFT algorithm and K-means algorithm in the process of feature extraction and cluster respectively.K-means algorithm,however,is heavily dependent on the initial value.The cluster result of K-means is easy to appear forced cluster when the class number is unknown.And SIFT algorithm is easy to cause data overflow and increase the retrieval time.Two novel feature extraction methods,called SIFT_CRONE and Color_HU respectively,were proposed and ISODATA algorithm was introduced in this paper.The SIFT_CRONE feature extraction method determines the key points of the image using SIFT algorithm,calculates the pixel gradient around the key points using CRONE operator and describes the key points by vector.Its advantages are that it keeps the advantages of SIFT features and reduces the time costs of retrieval.In Color_HU feature extraction method,we determined the key points and the effective area by SIFT,and calculated color histogram and HU moment of the effective area to reduce the feature dimension and the retrieval time costs.Meanwhile,we presented an adaptive parameter estimation algorithm for ISODATA.The experimental results show that the ISODATA algorithm can avoid the dependence on initial value of K-means,and can obtain ideal results when the cluster number is unknown.Two proposed feature extraction methods have their own advertages,and both can shorten the time of image retrieval and improve the retrieval efficiency.

关 键 词：词汇树 图像检索 K-MEANS ISODATA CRONE算子 SIFT特征 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]