验证和识别相融合的深度行人识别网络

找到能减小类内距离、增大类间距离的特征表示方法是行人识别的一个挑战。提出一种基于行人验证和识别相融合的深度网络模型来解决这一问题。首先，识别监督学习网络模型增加不同个人的类间间距，验证监督学习网络模型减少同一个行人的类内间距；然后，将行人验证和识别的深度网络融合，提取到更有分辨能力的行人特征向量；最后，采用了联合贝叶斯的行人比对方法，通过监督学习排名的方式，提高行人比对的准确率。实验结果表明，所提方法在VIPeR库上同其他深度网络相比有较高的识别准确率，融合网络与单独的识别和验证网络相比有更高的收敛速度和识别准确率。

Deep network for person identification based on joint identification-verification

It is a challenge for person identification to find an appropriate person feature representation method which can reduce intra-personal variations and enlarge inter-personal differences. A deep network for person identification based on joint identification-verification was proposed to solve this problem. First, the deep network model for identification was used to enlarge the inter-personal differences of different people while the verification model was used for reducing the intra-personal distance of the same person. Second, the discriminative feature vectors were extracted by sharing parameters and jointing deep networks of identification and verification. At last, the joint Bayesian algorithm was adopted to calculate the similarity of two persons, which improved the accuracy of pedestrian alignment. Experimental results prove that the proposed method has higher pedestrian recognition accuracy compared with some other state-of-art methods on VIPeR database; meanwhile, the joint identification-verification deep network has higher convergence speed and recognition accuracy than those of separated deep networks.