| Image is one of the main carriers of information transmission.Low-resolution images with missing details due to the constraints of hardware devices such as imaging systems and bandwidth constraints during image transmission can’t meet the needs of real life.Therefore,how to improve the image resolution to obtain high-quality images has become an urgent research issue.Image super-resolution reconstruction refers to the use of digital image processing technologies to improve the resolution of images without upgrading the hardware.These technologies have been widely applied in the fields of medical detection,communication,public safety,and remote sensing imaging.With the development of machine learning and neural networks related theoretical research,research on image superresolution reconstruction algorithms based on deep learning is becoming mainstream recently.This paper aims to super-resolution reconstruction technology based on generative adversarial networks and deep learning,proposes two novel image super-resolution reconstruction algorithms,which acquire higher objective evaluation indicators and better subjective visual effects.The research contributions of the paper are as followings.(1)A super-resolution reconstruction algorithm SRRGAN(Super-Resolution Relativistic Generative Adversarial Networks)based on an improved residual block and an enhanced discriminator for generative adversarial networks framework is proposed.Aiming at the problems of traditional super-resolution reconstruction algorithms based on residual networks with large parameters and high computational complexity,SRRGAN is designed to remove the redundant residual blocks of batch normalization layers to form generator network,which can speed up the training process of network.Aiming at the problems that the existing super-resolution models can’t restore the texture details of images well,SRRGAN introduces the relativistic average GAN theory into the loss function,and sets discriminator to predict the probability that the real high-resolution image is relatively more realistic than the reconstructed high-resolution image,which can enhance the discriminative ability of discriminator and provide more powerful supervision for model training.The experimental results prove that SRRGAN algorithm can better restore the detailed information of LR images and improve the qualities of reconstructed images.(2)A super-resolution reconstruction algorithm TESRGAN(Texture Enhanced SuperResolution Generative Adversarial Networks)based on generative adversarial networks combining with texture loss is proposed.Aiming at the problems that the traditional superresolution reconstruction algorithms based on generative adversarial networks is difficult to train and generate artifacts in reconstructed images,TESRGAN introduces the dense residual blocks based on the idea of dense convolutional networks to deepen the structure of generator network.TESRGAN uses VGG19 network as the basic framework for discriminator network to control the training direction of image generation.For the loss function,four losses are introduced to constitute the total objective function for generator.The content loss is used to ensure the consistency of low-frequency information between the reconstructed high-resolution images and the original high-resolution images.The adversarial loss is optimized using WGAN-GP theory,which can provide more powerful and effective supervision for network training.The perceptual loss is calculated by using the feature information before activation layer,which is helpful to restore more accurate brightness and realistic textures.Finally,the texture loss is introduced to encourage the matching of local texture details to obtain better results.The experimental results show that TESRGAN algorithm is superior to other compared algorithms in the objective evaluation indices without losing too much speed,and it also significantly improves the subjective visual evaluation such as brightness information and texture details. |
