Research And Design Of Image Super-Resolution Algorithm Based On Generative Adversarial Networks

With the rapid development of information technology,images have become an important methord for people to acquire,share and understand information.Moreover,pictures with higher quality can convey more content effectively.However,in real life,due to the impact of imaging equipment performance,imaging environment and other factors,people often get poor display quality images,these images can not provide necessary details,causing some trouble for researchers to carry out follow-up analysis and research.Therefore,how to improve the image display quality has become an urgent problem to be solved in academia and industry.Image super-resolution reconstruction technology is to convert one or more low resolution images into high resolution images through algorithm processing when the hardware cost is not significantly improved.At present,this technology has been widely used in medical,communication,military and other fields.With the development of deep learning technology,image super-resolution reconstruction algorithm based on deep learning rapidly replaces the traditional algorithm with better reconstruction effect and higher processing speed.Generative adversarial network is an important research direction in the field of deep learning and has attracted the attention of researchers because of its powerful data generation ability.This paper proposes two improved algorithms for image super-resolution reconstruction based on generative adversarial network.The specific work is as follows:(1)Aiming at the problems of poor visual effect and smooth details generated by image super-resolution reconstruction model based on convolutional neural network,an improved image super-resolution reconstruction model based on generative adversarial network is proposed.A generator with residual dense connections is designed as the basic nonlinear mapping unit,and introduce pyramid attention module to improve the generating network’s ability to generate high-resolution images.In the discriminator network,the relative average discriminator is used to improve the discriminator ability of the discriminant,and the spectral normalization convolution is used instead of BN to ensure the stability of the generated adversarial network training.The whole model is optimized by using perceptual loss.Experimental results show that the algorithm can recover details better and generate images with better visualization quality.(2)Aiming at the problem that the image super-resolution reconstruction algorithm based on the generative adversarial network is prone to structural distortion,a new image super-resolution reconstruction algorithm based on the gradient guidance and generative adversarial network is proposed.By introducing gradient branch to transmit image gradient information,the generator network ensures that the image edge structure does not distort.Morever,an improved multi-scale residual module is proposed for image branch and gradient branch,which makes it easier for the model to capture multi-scale feature information.The discriminator network uses WGAN-GP to enhance the gradient controllability and improve the stability of the generative adversarial network training.Experimental results show that the proposed algorithm can effectively alleviate the structural distortion of the existing image super-resolution reconstruction algorithms based on generative adversarial network,and the computational complexity of the model is lower than that of the existing image super-resolution reconstruction algorithm based on gradient guidance SPSR.