在无监督学习领域中生成对抗网络是近几年发展速度较快的一个研究方向，其主要特征是用一种概率估计的方式去逼近未知分布的模型。运用这种模型可以避免复杂的运算和编码，特别是在计算机视觉方面，可以生成质量较好的图片。本文基于循环生成对抗网络，训练一个风格迁移的神经网络，通过输入一张采样图片转化为输出一张与采样图片不同风格的图片。为防止生成器学习到具有欺骗性的虚假数据，本文采用加入一个新的生成器，把第一个生成器的输出当作输入再次使用，使生成器的输出和原图具有较高的相似性，同时不丢失原图片的特征并且确保输出一个和原始输入相似的图片。实验仿真数据集选取真实校园作为场景，在训练初期并不能较好的还原回原始图片，这意味着生成器使用了虚假的输出结果，当训练次数达到10,000次以上，结果显示可以较好的还原回原始图片，证明第一个生成器的输出保留了大量原始图片特征，输出结果较为可靠。
In the field of unsupervised learning, generative confrontation network is a research direction that has developed rapidly in recent years. Its main feature is to use a probability estimation method to approximate an unknown distribution model. Using this model can avoid complex calculations and coding, especially in computer vision, can generate better quality pictures. This article is based on the cyclic generative confrontation network, training a style transfer neural network, inputting a picture and outputting a picture of its different styles. In order to prevent the generator from learning deceptive false data, this article adopts adding a new generator, and uses the output of the first generator as input again, so that the output of the generator has a higher similarity with the original image, while not losing the characteristics of the original picture and ensuring that a picture similar to the original input is output. The experimental simulation data set selects the real campus as the scene, and the original picture cannot be restored well in the initial training stage. This means that the generator uses false output results. When the number of training times reaches 10,000 or more, the results show that it can be restored well. Going back to the original picture, it is proved that the output of the first generator retains a large number of original picture features, and the output result is relatively reliable.
