基于边界平衡生成式对抗网络的图像生成

基于边界平衡生成式对抗网络的图像生成(论文13000字)
摘要：随着人们对图像质量要求的不断提高，生成高分辨率清晰的图像已成为一项重要的研究课题。最新提出一种图像生成方法刷新了计算机生成图像的质量记录-----基于边界平衡生成式对抗网络的图像生成，新提出的方法解决传统生成式对抗网络训练复杂、控制样本多样性困难、平衡判别器和生成器困难等难题，最终在图像生成领域取得了卓越的成果，达到了目前最先进的水平。本文简单介绍了生成式对抗网络基本原理，着重分析了边界平衡生成式对抗网络系统，其中包含了平衡生成器与判别器新的方法，并提出了相对应的损失函数，根据Wasserstein距离设计出损失函数，Wasserstein距离是训练基于自编码器生成对抗网络，该方法不仅在训练期间实现了平衡，而且还提供了一种全新的近似收敛方法，达到快速稳定训练的目的，因此可以完成高质量图像生成的任务。本文还推导能够控制图像多样性和视觉感官质量之间平衡的方法。在本次实验过程中，主要专注于生成高质量图像，在高分辨率下，只需要相对简单的模型结构和标准的训练程序即可实现高质量图像生成。
关键词：边界平衡生成式对抗网络、生成器、判别器

Image Generation Based on Boundary Equilibrium Generative Adversarial Networks
Abstract:With the continuous improvement of image quality requirements, generating high resolution and clear images has become an important research topic. A new method of image generation is proposed, which refreshes the quality record of computer generated images. The new method resolves the problems of complex training, difficulty in controlling sample diversity, difficulty in balancing discriminator and generator of traditional generative countermeasure network. Finally, it achieves excellent results in the field of image generation. At present, the most advanced level. In this paper, the basic principle of generative countermeasure network is briefly introduced, and the boundary-balanced generative countermeasure network system is emphatically analyzed, which includes a new method of balancing generator and discriminator, and the corresponding loss function is proposed. According to Wasserstein distance, a loss function is designed (Wasserstein distance is a training-based self-encoder-generated countermeasure network). This method is not only in the training period, but also in the training period. It also provides a new approximate convergence method to achieve the goal of fast and stable training. Therefore, the task of high quality image generation can be accomplished. This paper also derives a method that can control the balance between image diversity and visual sensory quality. In this experiment, we mainly focus on generating high-quality images. In high resolution, we only need relatively simple model structure and standard training procedures to achieve high-quality image generation.
Key word:Boundary Equilibrium Generative Adversarial Networks、Generator、Discriminator

目录-----------------------------------------------------------------1
摘要-----------------------------------------------------------------2
Abstract-------------------------------------------------------------3
一绪论------------------------------------------------------------4
    1.1研究目的与意义-----------------------------------------------4
    1.2国内外研究现状-----------------------------------------------4
    1.3研究的主要工作-----------------------------------------------5
    1.4论文结构-----------------------------------------------------5
二 相关技术简介-------------------------------------------------5
2.1生成模型的研究-----------------------------------------------5
    2.2生成式对抗网络-----------------------------------------------6
2.2.1生成式对抗网络介绍-------------------------------------6
        2.2.2生成式网络模型的优点-----------------------------------7
        2.2.3损失函数介绍-------------------------------------------7
    2.3优化的边界平衡生成式对抗网络---------------------------------8
    2.4小结---------------------------------------------------------8
三 基于边界平衡生成式对抗网络的系统设计-----------------8
3.1系统模型------------------------------------------------------8
3.2边界平衡生成式对抗网络算法流程-------------------------------10
3.3具体工作-----------------------------------------------------10
3.3.1设计损失函数-------------------------------------------11
3.3.2平衡生成器与判别器-------------------------------------11   
3.3.3优化边界平衡生成式对抗网络-----------------------------12
3.4数据集、开发工具---------------------------------------------14
    3.4.1CelebA人脸图像数据集介绍-------------------------------14
3.4.2开发工具介绍-------------------------------------------14
3.5实验结果-----------------------------------------------------14
        3.5.1生成图像的多样性和质量---------------------------------14
        3.5.2生成图像的空间连续性-----------------------------------15
        3.5.3测试收敛度量和图像质量---------------------------------16
    3.5.4测试网络平衡效果---------------------------------------16
3.6小结---------------------------------------------------------17
四 总结与展望----------------------------------------------------17
4.1总结----------------------------------------------------------17
4.2展望----------------------------------------------------------17
参考文献-----------------------------------------------------------19
致谢----------------------------------------------------------------20