使用重建损失提升生成对抗网络性能

需积分: 50 99 浏览量更新于2024-07-14 收藏 128KB PDF 举报

"Improved Generative Adversarial Networks with Reconstruction Loss" 在生成对抗网络（Generative Adversarial Networks, GANs）的研究中，一种常见的问题是模式塌缩（mode collapse）和训练不稳定性，这些问题可能导致生成器无法捕捉数据的全部多样性。为了解决这些问题，研究人员提出了一种名为“具有重建损失的改进型生成对抗网络”的方法。这个创新的正则化方案旨在利用鉴别器（discriminator）的学习能力来改善生成器（generator）的表现。基本思路是引入一个重建损失（reconstruction loss），它要求生成器不仅生成逼真的新样本，还要能够根据鉴别器学到的特征重建真实的数据。这样，生成器将被引导去生成更接近实际数据分布的样本，从而避免模式塌缩。重建损失被添加到原始GAN的目标函数中，使得生成器在生成新样本时，同时考虑与鉴别器特征的匹配程度。重建损失的引入对GAN的性能有显著提升，特别是在多个数据集上的图像生成任务中，它能够生成质量更高的图像。此外，这种方法还展示了与其它正则化技术（如梯度惩罚）的兼容性，可以进一步优化不同GAN架构的性能。实验在标准的深度卷积生成对抗网络（Deep Convolutional GAN, DCGAN）和复杂的残差网络（ResNet）架构上进行，结果证实了该方法的有效性和鲁棒性。论文作者Yanchun Li、Nanfeng Xiao和Wanli Ouyang的这项工作是在2018年被接受发表在《神经计算》（Neurocomputing）期刊上的。他们的研究为GAN的训练稳定性和生成多样性提供了一个实用的解决方案，对于理解如何通过特征级别的反馈来指导生成器的训练具有重要的理论和实践价值。通过这种方式，GAN可以更好地模拟真实世界的数据分布，这对于各种应用，如图像生成、数据增强和半监督学习等，都具有巨大的潜力。

ACCEPTED MANUSCRIPT

matched the prior distribution. Another diﬀerence with our method is that

AGE augmented the objective function with reconstruction loss of the latent

codes. The adversarial loss of AGE needed to compute the mean and varia-90

tion of every sample, which leaded to slow training. Another main shortage of

AGE was the generated data of low quality. We conjecture that this may be

because the adversarial loss in AGE is between the posterior distribution of real

data/generated data and the prior distribution, rather than directly between

the real data distribution and the generated data distribution like in original95

GAN.

From the perspective of adopting generator to reconstruct real data, the re-

lated works include MR-GAN [27] and α-GAN [28]. Both MR-GAN and α-GAN

contained a generator, an encoder and two discriminators. Their generators had

the same goals as our generator, which were to produce real-like data and de-100

code training data. The diﬀerence between MR-GAN and α-GAN was that the

other three networks had diﬀerent objective functions. Speciﬁcally, the encoder

optimization signal of MR-GAN only came from the real data reconstruction

loss, while the encoder optimization signal of α-GAN included reconstruction

loss of real data and judgment loss of latent space discriminator. In MR-GAN,105

one discriminator separated real data and their reconstruction, and the other

discriminator separated the reconstruction of real data and the generated data.

In α-GAN, the data space discriminator diﬀerentiated training data, reconstruc-

tion data and the generation, while the latent space discriminator diﬀerentiated

whether the codes came from posterior or prior. Due to the need to optimize110

four networks simultaneously, both MR-GAN and α-GAN were time-consuming

with high computational complexity.

The adversarial training in the above mentioned approaches are not directly

between the generated data and training data like in standard GAN. The ad-

versary in our method is between training distribution and generated distribu-115

tion, and improves the performance with an additional regularization. From

this viewpoint, the related works are DFM, WGAN-GP, DRAGAN [25, 32, 34].

DFM augmented the objective function of the generator with a denoising fea-

剩余32页未读，继续阅读

weixin_38706747

粉丝: 5
资源: 962

使用重建损失提升生成对抗网络性能

生成对抗网络的详细介绍PPT

生成对抗网络原理及代码解析.pdf

基于循环生成对抗网络的超分辨率重建算法研究.docx

基于生成对抗网络的图像超分辨率重建

基于生成对抗网络的超分辨率重建算法

详细介绍，生成对抗网络（GAN）进行去雾的方法

基于生成对抗网络的图像压缩算法的定义

生成对抗网络研究现状

生成对抗网络与编码解码网络的区别

生成对抗网络常用领域

最新资源