一种基于GAN的手势图像生成方法

A Hand Gesture Images Generation Method Based on Generative Adversarial Network

对近年来两种主流图像生成算法:变分自编码器和对抗式生成网络进行研究。提出一种对抗式生成网络的手势图像生成方法,用以实现数据增强。采用对抗学习的方式,分别设计了卷积判别网络模型和反卷积生成网络模型,并使用自适应学习率的方式优化训练过程,根据美国手势语数据集ASL中的部分手势图片生成大量新的手势图像。设计一组混合使用真实手势图片与生成手势图片作为训练集的对照实验测试生成效果。实验结果表明,生成图片作为训练集能达到和真实图片相似的效果。

In this paper,two kinds of mainstream image generation algorithms in recent years are studied,which are Varia⁃tional Auto-Encoder(VAE)and Generative Adversarial Networks(GANs).A gesture image generation method based on GANs is proposed to implement gesture recognition training data argumentation.Using the method of adversarial studying,convolutional dis⁃criminative network and de-convolutional generative network are designed and trained by using adaptive learning rate.A large num⁃ber of new gesture image are generated from some of hand gesture pictures in American Sign Language datasets.A controlled experi⁃ment is designed to test the generation effect by using a mixture of real gesture images and gesture images.The results show that us⁃ing generated images can achieve similar effect as real images.