关键词: 词汇树,图像检索,K-means,ISODATA,CRONE算子,SIFT特征

Abstract: 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.

Key words: Vocabulary tree,Image retrieval,K-means,ISODATA,CRONE,SIFT