This paper is about super-resolution (SR) !! Really many many SR algorithms have been proposed in the literature. Before the success of convolutional neural network (CNN), sparse-representation based SR algorithms showed the best SR performance.
The introduction of CNN into computer vision feild is truely a big impact. CNN has replaced all the existing records not only in high level vision problem but also in low level vision problem.
Many CNNs have been proposed for single image SR, and they achieved great success on recovering high resolution images from low resolution images. However, they still have a limitation in recovering lossed high frequency compenents from the low resolution images, such as texture.
<This picture is from the original paper. If we use MSE-based loss, many pathces can be the solution>
Theoretically, it is impossible to recover the lossed high frequency components from the low resolution images. Use your imagination. Even a human cannot draw the true hair, which exists in the high resolution image but disappeared in the row resolution image, on the low resolution image. However, a human can draw hair on the row resolution image as if the hair was there.
The goal of this paper is to create a high resolution image that is photo-realistic. The created image may not be the same as the true high resolution image. However, people may say, with high probability, "the quality of this image is very good!!!"
In order to achieve this goal, the authors use a generative adversarial network (GAN). The basic of GAN is
1) Generator network produces a high resolution image from a low resolution image, which can make Discriminator network believe the high resolution image is an original.
2) Discriminator network discriminates if an input is an original high resolution image or the high resolution image created by Generator network.
I am not sure if the authors first propose using GAN concept for SR. The authors said that their main contributions are
1) Build a new ResNet for SR (this is used as Generator network), which shows the state of the art performance in term of PSNR
2) Propose perceptural loss fuction, which replaces the existing MSE based loss function
Here I show the representative result of the paper.
<This picture is from the original paper. If we use MSE-based loss, many pathces can be the solution>
What do you think of the result? What about you see the result on the display of big size and high resolution?
