基于双流卷积神经网络和生成式对抗网络的行人重识别算法  

Dual stream ConvNet-Gan for person re-identification

作　　者：林通 陈新[1] 唐晓[1] 贺玲[1] 李浩[1] Lin Tong;Chen Xin;Tang Xiao;He Ling;Li Hao(People′s Liberation Army Force Army Early Warning Academy,Wuhan 430019,China)

机构地区：[1]中国人民解放军空军预警学院,湖北武汉430019

出　　处：《信息技术与网络安全》2020年第6期7-12,共6页Information Technology and Network Security

基　　金：国家自然科学基金(61502522)。

摘　　要：近年来,针对行人重识别问题的深度学习技术研究取得了很大的进展。然而,在解决实际数据的特征样本不平衡问题时,效果仍然不理想。为了解决这一问题,设计了一个更有效的模型,该模型很好地解决了目标的不同姿态的干扰以及数据集中的图片数量不足的问题。首先,通过迁移姿态生成对抗网络生成行人不同姿势的图片,解决姿态干扰及图片数量不足的问题。然后利用两种不同的独立卷积神经网络提取图像特征,并将其结合得到综合特征。最后,利用提取的特征完成行人重识别。采用姿势转换方法对数据集进行扩展,有效地克服了由目标不同姿势引起的识别误差,识别错误率降低了6%。实验结果表明,该模型在Market-1501和DukeMTMC-Reid上达到了更好的识别准确度。在DukeMTMC-Reid数据集上测试时,Rank-1准确度增加到92.10%,m AP达到84.60%。The performance of current algorithm of person re-identification (ReID) has been made great progress, but there are still a lot of difficults in dealing with different backgrounds and poses. In order to cope with the difficult in actual scenes, this paper proposes a novel model, which can very well solve the problem that the characters have different poses and the number of pictures in the dataset is insufficient and uneven. Firstly the IDs with a relatively small number of images are screened out to expand by transferring pose and solve the problem of imbalanced datasets. Then two different convolutional neural networks are used to extract picture features and they are joined to get different characteristics. In the end, the task with extracted features is completed. This model solves the impact of different poses on the recognition effect very well. At the same time, it uses two independent feature extraction networks to extract the features of the per-son more comprehensively. Experimental results confirm that this model significantly improves performance and achieves high accuracy in Market-1501 and DukeMTMC-Reid. When tested on the DukeMTMC-Reid dataset, Rank-1 was in-creased to 92. 10 %, and mAP was increased to 84. 60 %.

关 键 词：行人重识别 卷积神经网络 生成式对抗网络 姿势迁移 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]