生成式对抗网络在手写字的性能分析

Performance Analysis of Generative Adversarial Network for Handwriting

针对深度卷积神经网络(DCNN,deep convolutional neural network)广泛应用在生成式对抗网络(GAN,generativeadversarial networks)中并取得优异的效果,为了探究不同结构的模型生成的图片质量各异的原因,实验对比GAN、DCGAN、WGAN-GP、InfoGAN等模型在手写字数据集(Mnist)的生成模型损失函数、判别模型损失函数以及图片质量。通过对比实验表明,InfoGAN通过计算数据之间的互信息影响生成新数据的方式,其生成式模型损失函数和判别式模型损失函数极值均小于1.0,远远小于其他模型,随着迭代次数的增加,损失函数曲线较为稳定,且图像生成质量更佳。

Deep Convolutional Neural Networks(DCNN,deep convolutional neural network)are widely used in Generative Adversarial Networks(GAN,x1x and have achieved excellent results.In order to explore the reasons for the different quality of images generated by models of different structures,Experiments compare GAN,DCGAN,WGAN-GP,Info GAN and other models in the handwriting data set(mnist)generation model loss function,discriminant model loss function and picture quality.Experiments show that Info GAN generates new data by calculating the mutual information between the data.The loss function of the generative model and the discriminant model loss function are less than 1.0,which is much smaller than other models.As the number of iterations increases.The loss function curve is more stable,and the image generation quality is better.