【摘要】 生成对抗网络(GAN)是目前图像生成领域中一种新的、有效的训练生成模型方法。深度卷积生成对抗网络(DCGAN)作为GAN的一种延伸,将卷积神经网络引入到生成模型中进行无监督训练。但DCGAN的线性卷积层对于下层数据块是一个广义线性模型,其抽象层次较低,生成的图像质量不高,并且在模型性能度量方面仅以主观的视觉感受来评判图像质量。针对以上问题,文中提出了一种多层感知器深度卷积生成对抗网络(MPDCGAN),采用多层感知器卷积层取代广义线性模型在输入数据上进行卷积,以捕获图像更深层次的特征,并采用定量评估方法Frechet Inception Distance(FID)衡量图像生成质量。在4种基准数据集上的实验结果表明,采用MPDCGAN生成的图像的FID值与图像质量呈负相关关系,且图像生成质量随着FID值的降低得到了进一步的提高。更多还原

【Abstract】 Generative adversarial network(GAN) is currently a new and effective method for training generative model in image generation.As an extension of GAN,deep convolutional generative adversarial network(DCGAN) introduces convolutional neural networks into the generative model for unsupervised learning.However,the linear convolutional layer of DCGAN is a generalized linear model for the underlying data block.The abstraction level of DCGAN is low and the quality of the generated image is not high.In terms of model performance metrics,image quality is judged only by subjective visual perception.Aiming at the above problems,multi-layer perceptron deep convolutional generative adversarial network(MPDCGAN) was proposed,and the multi-layer perceptron convolutional layer was used to replace the generalized linear model to convolve the input data to capture the deeper features of the image.In order to evaluate the quality of the generated image,a quantitative evaluation method named Frechet Inception Distance(FID) was used.The experimental results on the four benchmark datasets show that the FID value of the image generated by MPDGAN is negatively correlated with the image quality,and the image quality is further improved with the decrease of the FID value.更多还原