Papers connection: https://arxiv.org/abs/1804.02958v1
I. Introduction
Generated using image compression against the network, in fact, equivalent to the original place of the decoder by a generator. The picture decoder restored to the original encoded image, generated by the encoder is encoding, it generates a codeword length and quality is directly related to the image, and this may limit the coding rate is further reduced. In this paper, the authors use the network as a decoder to generate confrontation is to solve this problem. When the encoding process, not the whole image is encoded, but only a certain part of the encoded and then restore the original image restored by decoding the coding part, the coding part is not automatically generated by the generator G, so that only part of the picture needs to be encoded, can greatly improve the compression ratio.
II. Network Structure
1.Global generative compression
It includes two specific network structure, a first called the Global generative compression (GC), this method is applicable to the whole image to be saved. Which part needs to be saved, which part needs to be generated by the network is automatically selected according to their semantic graphs and optimization objectives.
Here the objective function comprises three parts, the first two equations are GAN objective function, the third control expression is generated with respect to the original picture image distortion last expression is to control the compression ratio, can be adjusted by the β size to adjust the compression rate.
2.Selective generative compression
A second structure is called Selective generative compression (SC), this structure is generally used under certain scenarios, such as video calls, people tend to pay more attention to the video, while the background and do not care. Therefore, only the portion encoding the portrait and the background portion generated by the generator automatically. Coding for which part, which part is generated is by a binary pattern control, to be generated as part of the value 0, the value part is a need to save.
There are two ways of training, one is randomly selected 25% of each of the training images to save, to generate the rest; the other is to set a fixed size of the window, the window internal storage, an outer portion of the window generation. SC using the objective function and substantially the same GC, except that during training, the third part of the objective function, only the area to be saved is calculated, since it has been postulated that this portion is not important.
III. Evaluation Criteria
In particular, when the compression rate is small when measured by PSNR and SSIM picture quality has no meaning. Because to PSNR, for example, it is more concerned about how much local information is lost, and in the case of the compression ratio approaches 0, the image distortion is already very large, people are more concerned about the overall image changes, rather than local information is lost, so this time with a little image quality measure PSNR sense. So authors used mIoU to estimate the quality of the picture, the authors compared the difference image after image and the original image is compressed semantic segmentation obtained.
In addition, the authors also by way of a user survey to verify with a better visual effect by compressing the images obtained in this way.
