Abstract: In the era of big data, fast and accurate indexing algorithm is crucial for the information retrieval. As the search performance of Single Table Hash Retrieval（STHR） method utilizing random projection is too low, a novel multi table image hash retrieval method based on principal component is proposed. In order to get efficient hash encoding to distinguish the feature of different semantic samples, discriminative image features are holden by the principal component analysis method. Furthermore, hash index tables are established by using Feature Clustering as a guide for learning hash projection. Meaning while, the orthogonal rotation matrix is employed to optimize the hash projection, which ensures that the samples with the same semantics have similar hash codes. Finally, a series of experiments are conducted on CIFAR-10 and Caltech-256 data sets comparing with the related methods. The results show that the proposed method gains a great improvement in terms of retrieval performance.

Key words: distinguishing feature of image, Principal Component Analysis（PCA）, feature clustering, orthogonal rotation matrix, hash function, multi table index

摘要： 大数据时代的到来，快速而准确的索引算法对信息检索至关重要。针对基于随机投影构成的单表哈希检索方法导致搜索性能低的问题，提出一种基于主成分的多表图像哈希检索方法。为了得到高效的哈希编码保证不同语义样本特征的区分性，首先通过主元分析方法保留训练集具有区分性图像特征，此外利用特征聚类作为学习哈希投影的指引构建多个索引表；其次采用正交旋转矩阵对哈希投影进行优化，保证了相同语义的样本具有相似的哈希码。最后分别在CIFAR-10和Caltech-256数据集上与相关方法进行比较，实验结果表明提出的方法提高了检索性能。

关键词: 区分性图像特征, 主元分析, 特征聚类, 正交旋转矩阵, 哈希函数, 多表索引