Research And Application Of Super Resolution Image Reconstruction Algorithm Based On Generative Adversarial Networks

Image super-resolution reconstruction has a wide range of applications in the field of national defense,military and medical imaging,which is one of the research hotspots in the field of computer vision in recent years.Image super-resolution reconstruction refers to the transformation from single or multiple low-resolution images to high-resolution images.How to improve the resolution of reconstructed image is an urgent problem.In order to solve the above problems,this paper applies the Generative Adversarial Networks(GAN)to the task of image super-resolution reconstruction,and proposes two improved algorithms.Firstly,this paper introduces the research status of image super-resolution reconstruction technology and the theoretical basis of GAN,then proposes a super-resolution reconstruction algorithm based on GAN and applies it to the task of image reconstruction.The main contents of this paper are as follows:(1)Activate function adaption Generative Adversarial NetworksIn view of the mode collapse in the GAN,this paper designs an Activate function adaption Generative Adversarial Networks algorithm based on the activation unit.The nonlinear transformation of network parameters by activation function can increase the diversity of network parameters.However,most of the activation functions have no gradient in the negative direction of the network parameters,which will result in the information carried by the negative network parameters being discarded.The loss of parameter information will result in the network unable to learn more effective information from the training,and the reconstructed image can only restore part of the original data.However,the Activate function adaption Generative Adversarial Networks algorithm uses the Parametric Rectified Linear Unit(PRe LU)activation function when dealing with negative parameters,which carries adaptive factors and has continuous and lasting gradient values in the negative direction,so a lot of effective information is reserved for network training.The experimental results show that the Activate function adaption Generative Adversarial Networks algorithm has better performance in image reconstruction,and has better effect in objective and subjective evaluation indexes.(2)Triple Chain Train Generative Adversarial NetworksIn order to solve the problem that the learning rate of the generator and the discriminator does not match easily,which leads to the collapse of the pattern,this paper designs a Triple Chain Train Generative Adversarial Networks algorithm.When the learning rate of the generator and the discriminator does not match,the generator can not learn the effective information from the network training,and even lead to the interruption of network learning,resulting in the serious phenomenon of pattern collapse.Based on the original Wasserstein Generative Adversarial Networks(WGAN)model,the algorithm redesigns the network.Generator 1 and generator 2 will generate different parameter distributions in training.After that,we share the parameter distribution and learning experience with each other,which makes the generator network have higher reconstruction performance.The generated image data has more abundant information content and clearer image texture.Experiments show that this algorithm is superior to other representative models in objective and subjective evaluation criteria.