Research And Implementation Of Generative Adversarial Network Based On Human-computer Interaction

Generative Adversarial Nets(GAN)are image generation models in which generators and discriminators compete against each other to produce output.In this paper,an open human-computer interaction method is proposed,in which human subjective evaluation is added to the GAN training process to improve the quality of generated images and the personalized effect.The main research contents of this paper are as follows:1.The principle,application field and related models of generative adversarial network(GAN)are studied and analyzed.First,the principle of GAN is analyzed,and the loss and structure characteristics of GAN used in different tasks are summarized.Secondly,the application of image domain transformation,image super resolution,data enhancement and style fusion in GAN is summarized.Considering the important position of image domain transformation in GAN,the classical model and techniques of image domain transformation are emphically discussed.Finally,the Cycle GAN model for unsupervised domain transformation is studied and analyzed.Model PGGAN for high resolution image generation;The model MUNIT is used for a single model to achieve multimodal output and the model Sin GAN used a single image generation.2.In order to improve the generation quality of low-resolution images,a single-scale generation model S＿Open GAN is designed in this paper.S＿Open GAN contains an encoder,a learning network,and a decoder.The encoder includes three subsampled convolution kernels(all of which are 3×3 convolution).The learning network consists of three improved dense blocks(each dense block contains four 3×3 convolutions and four 1×1 convolutions for reducing parameters number),and it used IN normalization;The decoder consists of three upsampled convolutions(all of which are 3×3 convolution).S＿Open GAN uses the human-computer interaction method.The evaluator subjectively scores the generated image(0～1),and the score is fed back to the penalty item in the generator loss function to complete the generator parameter update.Experimental results on low-resolution datasets LSUN and Horese2 Zebra showed that the IS index(quality index)of S＿Open GAN increased by 51% on average compared with that of S＿Open GAN without interaction.Compared with Cycle GAN and DCGAN,the IS value of S＿Open GAN increased by 47% on average,and the LPIPS decrease by 52% on average.3.In order to improve the quality of high-resolution image generation and the level of personalization,this paper designs a multi-scale generation model M＿Open GAN,which is composed of four sub-generators with the same structure.Each sub-generator contains an encoder,a learning network and a decoder,and each sub-generator performs output upsampling respectively to complete the multi-scale generation task.The encoder consists of three convolutions(all of which are 3×3 convolution kernels).The learning network consists of eight improved residual blocks(each of which contains four 3×3 convolutions of the same size),in which each residual block uses Ada IN normalization.The decoder consists of three convolutions(all of which are 3×3 convolutional kernels)and two up-sampling blocks.M＿Open GAN also used a human-computer interaction approach.The experimental results show that: Compared with U-GAT-IT,MUNIT and Cartoon GAN,M＿Open GAN’s IS values on high-resolution dataset GTA,ADE20 K and Cityscape increased by 61%,KID decrease by44%,and LPIPS decreased by 49%,respectively.In addition,M＿Open GAN can achieve image personalization.