Real-time face identification via CNN and boosted hashing forest

This paper presents a new approach to constructing a biometric template using a Convolutional Neural Network (CNN) with Hashing Forest. The approach consists of several steps: training a convolutional neural network, transforming it to a multiple convolution architecture, and finally learning the output hashing transform via a new Boosted Hashing Forest technique. This technique generalizes the Boosted SSC (Similarity Sensitive Coding) approach for hashing learning with joint optimization of face verification and identification. The proposed network via hashing forest is trained on the CASIA-WebFace dataset and evaluated on the LFW dataset. The result of coding the output of a single CNN is 97% on LFW. For Hamming embedding, the proposed approach enables a 200 bit (25 byte) code to be constructed with a 96.3% verification accuracy and a 2000-bit code with a 98.14% verification accuracy on LFW. The convolutional network with hashing forest with 2000x7-bit hashing trees achieves 93% rank-1 on LFW relative to the basic convolutional network's 89.9% rank-1. The proposed approach generates templates at the rate of 40+ fps with a GPU Core i7 and 120+ fps with a GPU GeForce GTX 650.