Deep CNN based binary hash video representations for face retrieval

In this paper, a novel deep convolutional neural network is proposed to learn discriminative binary hash video representations for face retrieval. The network integrates face feature extractor and hash functions into a unified optimization framework to make the two components be as compatible as possible. In order to achieve better initializations for the optimization, the low-rank discriminative binary hashing method is introduced to pre-learn the hash functions of the network during the training procedure. The input to the network is a face frame, and the output is the corresponding binary hash frame representation. Frame representations of a face video shot are fused by hard voting to generate the binary hash video representation. Each bit in the binary representation of frame/video describes the presence or absence of a face attribute, which makes it possible to retrieve faces among both the image and video domains. Extensive experiments are conducted on two challenging TV-Series datasets, and the excellent performance demonstrates the effectiveness of the proposed network.