本文在比较多种仿射不变性区域的基础上，选取最大稳定极值区域(MSER)对图像内容进行分割和提取，并通过构造仿射不变量的方法，对提取出来的区域进行规整化，进而从规整化的区域中提取SIFT描述子，然后根据匹配目标的需要进行特征的组合、变换，以形成易于匹配、稳定性好的特征向量，从而把图像匹配问题转化为特征的匹配问题 通过改进的K均值算法对特征向量进行聚类，聚类中心作为图像的视觉关键词:利用视觉关键词的思想，把文本检索技术领域的方法移植到图像检索应用上；并研究视觉关键词之间的空间约束关系。-Most relevant between images often exist perspective, dimension, rotating change and fuzzy, partial sheltering and complex background general affine transformation, and in many image processing and application needs a stable and can in different image can repeat detection feature to eliminate the influence of these changes, so it has good stability and repeatable degrees and can match the local invariant features become sex research hotspot.