基于生成式对抗网络的小样本图像生成

Based on Generative Adversarial Networks

研究基于小规模数据集(即图像尺寸维度远大于样本集的数据规模)的图像生成问题,数据集规模过小会影响深度学习中生成式模型(Generative Models)生成图像的效果。针对小规模数据集图像生成问题,对已有深度卷积生成式对抗网络DCGAN进行了改进,提出了改进的MDCGAN(Modified DCGAN)。MDCGAN模型采用卷积层取代全链接层,采用带步长的卷积运算取代上采样运算。同时将条件信息y引入判别器和生成器中,条件信息y能够为生成式对抗网络增加条件,对生成数据起到监督作用。通过手写数字生成实验和建筑物轮廓生成实验证明,提出的MDCGAN能够基于小规模数据集生成高清晰度高逼真度的图像。

This paper studied the problem of image generation based on small-scale data sets(that is, the size of the image size is much larger than the data size of the sample set). Too small a data set size will affect the image generation effect of generative models in deep learning. Aiming at the problem of image generation on small-scale datasets, this paper improved the existing DCGAN and proposed an improved MDCGAN. Compared with the original GAN, MDCGAN uses a convolution layer instead of a full link layer, and uses convolution with a step size instead of up-sampling. At the same time, we introduced the condition information y into the discriminative model and the generation model. The condition information y can add conditions to the generative adversarial network and guide the data generation process. Experiments on handwritten digital generation and building contour generation prove that our proposed MDCGAN can generate high-resolution and high-fidelity images based on small-scale data sets.