后验概率改进Fisher向量的高性能图像检索算法  

An image retrieval method combing with texture classification and modified Fisher vector

作　　者：邹利华[1] 战荫伟[2] 

机构地区：[1]东莞职业技术学院计算机工程系,广东东莞523808 [2]广东工业大学计算机学院,广东广州510006

出　　处：《电子技术应用》2017年第10期119-123,共5页Application of Electronic Technique

基　　金：国家自然科学基金资助项目(60572078)

摘　　要：提出了一种高性能的图像检索方法,结合纹理分类和改进的Fisher向量实现图像检索。首先,将图像划分为互不重叠的图像子块,对每一图像子块依据纹理复杂度进行分类,对不同类别的图像子块提取不同的特征。其次,采用基于后验概率改进的Fisher向量进行特征编码,依据乘积量化和非对称距离计算方法,分段计算两特征向量之间的距离,快速求取相似度指标,据此进行图像检索。在Holidays数据集上进行图像检索的实验结果表明,该方法的查准率和召回率高,且耗费的查询时间少。An image retrieval method with high performance is proposed, which combing with texture classification and modified Fisher vector to realize image retrieval. First, the image is divided into non-overlapping image sub-blocks, each image sub-block is classified according to complexity of texture, and different features are extracted for different classes of image sub-blocks. Second, features are encoded by modified Fisher vector based on posterior probability, and the distance between two feature vectors is calculated segmentally according to product quantification and asymmetric distance calculation method, for rapid computing a similari- ty index and executing image retrieval. Experimental results for image retrieval on Holidays dataset show that, this method has high precision and recall, and less query time consuming.

关 键 词：图像检索 尺度不变特征变换 Fisher向量 高斯混合模型 灰度共生矩阵 

分 类 号：TN271[电子电信—物理电子学] TP391[自动化与计算机技术—计算机应用技术]