Research On Optimization Of Generative Confrontation Network And Image Translation Algorithm

As the mainstream information carrier of social networks,images contain rich information,can share daily life,record events,and convey feelings for people,and are widely used in people's daily life.Image translation is the mapping between the source domain and the target domain when the training samples are sufficient,which can convert one representation of the target into another.Image processing and computer vision problems can be regarded as an image translation problem,such as image deblurring,face makeup,style transfer,image restoration,and so on.The Generative Adversarial Network(GAN)proposed in 2014 is outstanding in the field of image translation,and can generate images with different contents according to needs.Image translation based on GAN network,the input is random noise,and it is difficult to quickly generate the required target content due to large degree of freedom,which is prone to oscillation to fail to reach equilibrium and partial mode collapse.As a result,the generated image features such as outline and texture are lost,resulting in poor image translation effect.In view of these shortcomings,this paper proposes to improve the U-Net image generation algorithm of the adversarial network model by changing the number of network jump connections,and conducts supervised image translation and Pix2 Pix generated adversarial network model comparison experiments,unsupervised image translation and Cycle GAN model comparison experiments.The improved model proposed in the paper can solve the problems of background distortion and lack of diversity in generated images.Firstly,the effects of Pix2 Pix generation against network optimization algorithms,learning rate,and number of iterations on the effect of image translation are studied.When the number of iterations is 200,the optimization algorithm is Adam,and the learning rate is 0.001,the network image translation effect is the best.Secondly,to solve the problem of poor image translation effect caused by the loss of contour and texture features of the image generated during the image translation process of the Pix2 Pix generation adversarial network,it is proposed to improve the U-Net generation adversarial network by changing the number of jump connections in the network Model image translation algorithm.The i-layer and the n-i layer of the U-Net model carry similar information,and the characteristics of the i-layer can be copied to the n-i layer through jump connections,making the generated image closer to the real image.The characteristics of the background,texture,and contours of images that are easily lost by a single jump connection.In the experiment,the image translation experiment is carried out by increasing the number of jump connections between the i-layer and the n-i layer to determine the optimal value of the number of network jump connections.Using Pix2 Pix to generative adversarial network and an improved U-Net model for experiments on the CUFS face database,the experimental results show that when the number of network jump connections is equal to 5,the image translation effect is optimal,and the MOS score of the image translation result is 4.2,PSNR reached 14.3634,SSIM reached 0.5763,L1 loss was 29292746,Cosin was 0.9732,higher than Pix2 Pix generated adversarial network model.Finally,the unsupervised learning method is adopted,the database is the horse2 zebra and apple2 orange libraries,and the improved network jump connection model is used to perform image translation experiments under unpaired data.The experimental results show that the improved network jump connection method has further improved the quality and diversity of the generated image compared to the CycleGAN model.The Fréchet Inception distance(FID)evaluation of the experimental generation results reached 166.39,which was 2.84 lower than the CycleGAN model solve the problem of background distortion of generated images,and the model convergence speed is fast.