Research On Image Super-Resolution Generation Countermeasure Model In Complex Environment

Image super-resolution is an important branch of computer vision.It has important research significance and application value in security,medical,military and other fields.In recent years,with the continuous development of neural network and other technologies,image super-resolution has made better achievements in accuracy and generation speed.In the process of image super-resolution training,the real training data is generally difficult to obtain.Based on this,most researchers use the artificially constructed training data for training.However,there is a problem in the image super-resolution model trained with artificially constructed data,that is,it has a good effect on artificially constructed low-resolution images and a poor effect on real low-resolution images.The main reason for this problem is that most of the artificially constructed low-resolution images are synthesized by interpolation,while most of the real low-resolution images are formed for complex reasons,which may be the result of the joint action of many influencing factors.In view of the problems existing in image super-resolution,we mainly study and improve two problems.The first is the image super-resolution under low illumination,and the second is the image super-resolution under few training samples.The specific research is as follows:Firstly,aiming at the image with low illumination and low resolution,this article proposes an image super-resolution generation countermeasure network model(LSRGAN)integrating illumination loss.The model inputs the LR-HR image pair into the model by constructing the low illumination low resolution high resolution image pair,uses the generator network to generate the generated samples similar to the real HR,and uses the discriminator network to distinguish whether the samples are real samples or generated samples,so as to achieve a better super-resolution effect in the case of low illumination and low resolution.Second: for the image super-resolution with few training samples,this article proposes an image super-resolution generation countermeasure network(MSRGAN)based on MAML.The model initializes the SRGAN model by introducing the MAML method in meta learning.Different image super-resolution scenes correspond to different tasks of MAML.After MAML training,the model gets a better initialization value,which can get a good image super-resolution effect by training fewer rounds with fewer training samples.At the same time,MSRGAN also has a good performance in dealing with the problem of image super-resolution in new scenes.