Current location - Education and Training Encyclopedia - Graduation thesis - Super resolution reconstruction
Super resolution reconstruction
SRGAN is a super-resolution paper of CVPR 2065 438+07, which brings the effect of super-resolution to a new height. The so-called super-resolution reconstruction is to restore a low-resolution image to a corresponding high-resolution image. Because there is a lot of information missing in the ground resolution image, it is a morbid problem to find the inverse solution, especially when restoring the high resolution image. Traditional methods restore high-resolution images by adding some prior information, such as interpolation, sparse learning and random forest based on regression method. CNN has achieved very good results in super resolution.

SRGAN is based on CNN and trained by GAN method to realize super-resolution image reconstruction. It contains a generator and a discriminator. The discriminator body is VGG 19, and the generator body is a series of remaining blocks. At the same time, a sub-pixel module is added behind the model, which draws lessons from the idea of sub-pixel network of Shi et al., so that the resolution of the picture is improved at the last network layer, the separation rate is improved and the calculation amount is reduced. The network structure given in this paper is as shown in the figure:

The loss function forms of generator and discriminator are also given:

1. The loss function of the generator is:

Among them, it is the perceptual loss function proposed in this paper.

Content is missing:; When training the network, a higher peak signal-to-noise ratio can be obtained by using the mean square error loss. In the general super-resolution reconstruction method, MSLOSS is used for content loss, but the image restored by MSLOSS will lose many high-frequency details. Therefore, in this paper, the generated image and the target image are input into the VGG network respectively, and then the Euclidean distance of the feature map obtained by them after VGG is calculated as VGG loss.

Anti-loss: In order to prevent the gradient of the generator from disappearing when the discriminator is trained, this paper will reduce the loss function of the generator. ? It has been revised.

2. The loss function of the frequency discriminator is:

It is similar to the loss function of ordinary generation countermeasure network discriminator.