基于Hadamard变换的高维图像检索方法

A New Indexing Method for High-Dimensional Image Databases Using Hadamard Transform

传统索引方法对高维数据进行近邻搜索时会面临维数灾难问题,向量近似方法是一种有效的高维检索方法。提出一种 Hadamard 变换域上的向量近似方法,在变换域能量最大的分量上建立顺序索引,然后建立近似向量文件。同时提出低維过滤算法,可以在近邻搜索过程中高效排除不匹配近似向量,减少 I/O 访问时间,提高查询效率。在大型高维图像特征库上的实验表明,该方法性能优于小波变换域的向量近似方法。

Traditional indexing methods face the difficulty of ＇ curse of dimensionality＇ at high dimensionality. The vector approximation file（VA-File）approach based on wavelet transform is a very efficient high dimensional indexing method. In this paper, a new VA-File approach in the Hadamard transform domain is introduced. This approach combines Hadamard transform and principle component filtering algorithm, which can reduce the searching complexity and I/O cost dramatically on large image databases. Experiment results show that the new method is more efficient than VA-File based on wavelet transform.