基于CNN和DLTL的步态虚拟样本生成方法

针对步态识别在反恐、安防领域亟待解决的小样本问题，提出了一种基于深度卷积神经网络（convolutional and neural network，CNN）和DLTL（dual learning and transfer learning）的步态虚拟样本生成方法。首先用基于VGG19的深度卷积神经网络模型低层响应提取步态风格特征图，然后利用基于对抗网络的对偶学习（dual learning，DL）对风格特征图进行风格训练，得到风格特征模型；其次利用VGG19模型的高层响应提取步态内容特征图，然后让步态内容特征图对风格特征模型中的风格特征进行学习；最后使用迁移学习TL（transfer learning）获得步态虚拟偏移样本。实验结果表明，经过DLTL风格学习生成的步态虚拟样本虽然整体风格发生改变，但人体步态特征没有改变，可有效扩充小样本容量；当虚拟样本增加到一定数量时，步态识别率有所提升。该方法与现有步态虚拟样本生成方法进行对比实验，结果表明该算法优于现有方法，能够大量生成虚拟样本且稳定提高步态识别的识别率。

Gait virtual sample generation method based on CNN and DLTL

To solve the problem of small sample of gait recognition in the field of counterterrorism and security issues, this paper proposed a novel gait virtual sample generation method based on deep CNN and DLTL. Firstly, it extracted gait style feature map based on low-level of CNN model VGG19, and then it used the DL to carry on the style feature training. Thus it made style feature model. Moreover, high-level of VGG19 extracted gait context feature map, and then it used the TL to make context feature map carry on the style characteristic learning. Finally, it obtained the virtual migration samples. Experimental results demonstrate that these virtual samples remained individual gait feature but style feature. So this method can effectively expand small sample size. At the same time, when the number of virtual samples increase to a certain number, gait recognition rate has improved. This paper compared the method with the existing virtual sample generation method. The result shows that the method has a better performance, which can generate virtual samples in large numbers and improve the recognition rate of gait recognition steadily.