A Distributed Face Retrieval System for Large-Scale Social Networking Avatars

摘要

In order to solve the problems of poor accuracy and slow speed in face retrieval for large-scale social networking avatars, this paper proposes a real-time distributed face retrieval system supporting hundreds of millions for face images. Firstly, we detect face and learn the face feature vector by combined convolutional neural network models. Secondly, the locality sensitive hash algorithm is used to transform the high-dimensional feature into low-dimensional. Meanwhile, a distributed face feature vector database is proposed to balance the storage pressure for one single server node as well as to accelerate query speed by computing similarity score in parallel and merging sort results from all distributed server nodes. Experimental results show that the distributed system designed in this paper can efficiently, accurately and timely retrieve faces from large-scale social networking avatars.