基于空间上下文加权词汇树的图像检索方法

摘要

基于词汇树的词袋模型( Bag-of-Words)表示算法是目前图像检索领域中的主流算法。针对传统词汇树方法中空间上下文信息缺失的问题，提出一种基于空间上下文加权词汇树的图像检索方法。该方法在词汇树框架下，首先生成SIFT点的空间上下文信息描述。然后利用SIFT点间的空间上下文相似度对SIFT间的匹配得分进行加权，得到图像间的相似度。最后，通过相似度排序完成图像检索。实验结果表明，该方法能够大幅度提高图像检索的性能，同时，对大规模图像库有较好的适用性。%Vocabulary tree based Bag-of-Words ( BoW ) representation becomes popular for image retrieval recently. Aiming at the absence of spatial context information in conventional vocabulary tree approaches, an image retrieval approach using spatial context weighting based vocabulary tree is proposed. Within the framework of vocabulary tree, this approach firstly describes the spatial context information of SIFT features. Then, the matching scores between SIFT features are weighted based on spatial context similarity, and similarities between images are achieved. Finally, image retrieval results are obtained according to the ranking of similarities. The experimental results indicate that the retrieval performance is improved and the proposed approach applies to large scale databases.