目的:合成(synthesis)高分辨率图像
方法:提出一种新的变种条件标签GAN
贡献:提出ACGAN,提出新的生成图片质量(可判别性和多样性 discriminability and diversity of samples)评估方法Inception Accuracy和MS-SSIM,得出结论:高分辨图片在ACGAN下可以更为准确地做classification(Across 1000 ImageNet classes, 128 _ 128 samples are more than twice as discriminable as artificially resized 32 _ 32 samples. In addition, 84.7% of the classes have samples exhibiting diversity comparable to real ImageNet data.1)。
核心:In this work we demonstrate that that adding more structure to the GAN latent space along with a specialized cost function results in higher quality samples.
模型架构:
Loss函数:
The objective function has two parts: the log-likelihood of the correct source, LS, and the log-likelihood of the correct class, LC.
D is trained to maximize LS + LC while G is trained to maximize LC -LS.
Inception accuracy:
OpenAI提出Inception Score用以评价合成图像的真假,本文作者认为在生成图像不符合人的视觉感知时,Inception Score依然可能会很高,所以评判不准确。于是,作者将合成的图像输入训练好的Inception V3模型,对比模型的分类准确率。
不同分辨率的生成图片的Inception accuracy对比,人工降低分辨率后,accuracy也随之降低,说明生成的高分辨率图片不仅含有像素信息,还有一些帮助提高accuracy的信息。
总结:ACGan立足添加标签约束,以提高(高分辨率)图片生成质量,并提出新的生成图片质量和模式坍塌衡量标准。作者在附录中阐明了标签数量对模型稳定性的影响,实验中训练了100个AC-GAN,每个对10个标签进行分类。本文没有过多与其他现有的生成模型进行比较测试。