Photo-realistic single Image super-resolution Using a generative adversarial Network
2016.10.23
Summary :
contributions:
Gans provides a powerful framework to produce high-quality plausible-looking natural images. This article provides a very deep ResNet architure, using the concept of Gans, to form a perceptual loss function to close human perception to do photo-realistic Sisr.
The main contribution is:
1. For image SR, we have achieved new top-notch results, reducing the resolution by 4 times times, measured by: PSNR and Structure Similarity (SSIM). Specifically, we first use the Fast feature learning in LR space and batch-normalization to train the residual network.
2. A combination of content loss and adversarial loss as our perceptual loss is presented.
Method:
First, there are several concepts:
Super solved image $I _{sr}$: W * H * C; Low-resolution input image $I _{lr}$: RW * RH * C; High-resolution image $I _{hr}$: RW * RH * C.
Our ultimate goal is to train a production function, G, to predict the HR portion of the given input image LR input image. We achieve this by training a generator network as a feed-forward CNN $G _{\theta_{g}}$ parameter for $\theta_{g}$, here $\theta_{g} = {W_{1:l}; b _{1:l}}$ represents an L-layer deep network of weights and biases, and is obtained by optimizing a sr-specific loss function $l ^{sr}$. For a given training image $I ^{hr_{n}}$, n = 1,..., n corresponds to a low-resolution image: $I ^{lr}_n$, we optimize the following problem:
1. adversarial Network Architecture
The training and learning goal of the production-type confrontation network is a Minmax problem:
The author also considers the image super-resolution as such a process. A super-resolution image is produced by generator, making discriminator difficult to distinguish.
Is the approximate process involved in this article.
2. Perceptual Loss Function
The perceptual loss function designed in this paper is the guarantee of the performance of the algorithm.
Paper read: photo-realistic single Image super-resolution Using a generative adversarial Network