此学习笔记系列 is focus on a paper called "Generative Adversial Networks for Extreme Learned Image Compression"
文章连接: https://arxiv.org/abs/1804.02958
0. Abstract
整篇文章提出了一个以GANs framework 为基础的deep image compression model which includes an encoder, a decoder/generator and a multi-scale discriminator.
"A user study confirms that for low bitrates, our approach is perferred to state-of-the-art methods, even when they use more than double the bits." 重点在于即使之前的方法用了多于2倍的bits,在深层压缩之后,所解压的图片也没有此方法创造出来的图片效果好!
1. Introduction
1. modern engineered codecs (现代工程编解码器): WebP, JPEG2000, BPG. (最基本的方法)
2. Image compression systems based on deep nerual netowrks (DNNs) (进阶的方法)
Advantage:
Not only achieving competitive compression rates on