一种支持海量人脸图片快速检索的索引结构

An Index Structure of Fast Retrieval for Massive Face Image

为了在大规模的人脸数据库中准确快速地检索到所需图像,提出一种相似人脸检索方法。提取人脸图片的局部二值模式特征,通过建立投影矩阵将特征从欧几里德空间映射到汉明空间实现降维,再采用改进的多比特编码方法对降维后的特征进行编码,并生成图片签名,以曼哈顿距离取代汉明距离衡量签名之间的相似度,根据图片签名集合构建倒排索引表,通过倒排索引表高效地查找相似图片。包含20万张人脸图片的实验数据集的结果表明,该方法在保证检索精度的前提下,检索时间控制在0.15s以内,能够满足海量人脸图片检索的准确性与实时性要求。

In order to quickly and correctly retrieve the desired image from massive face image database,this paper proposes an efficient fast method for similar face image retrieval. It extracts Local Binary Pattern（LBP）features of face images and does dimensionality reduction by mapping the features from Euclidean space into Hamming space. A signature for each image is constructed by encoding dimensionality reduced features,using enhanced multi-bit encoding method.The similarity between each signature is judged by Manhattan distance instead of Hamming distance. It constructs inverted indexes from image signatures and fast retrieval is accomplished by using efficient inverted indexes. Experimental results on dataset containing 200 000 face images show that the retrieval time is less than 0.15 s,which satisfies the retrieval precision as well as the accuracy and real-time of large-scale face image retrieval.