A Dynamic Study On The Background Characteristics Of CycleGAN In Style Transfer

Image style conversion has a wide range of applications and high practical value.With the rapid development of high-performance computers and the introduction of deep learning research methods,nowadays,the society has higher requirements for the realization of image style conversion.Therefore,highquality image style conversion has become a research hotspot and difficulty in the field of deep learning.Cyclic consistent generation network(Cyle GAN)can use unpaired images to complete style conversion,which has achieved good results in some application problems.However,when the migration task involves distinguishing the foreground and background of the image,the effect of image generation is not satisfactory.For example,when using Cyle GAN for gender migration,the complex and changeable background of the image increases the uncertainty of migration,which causes obvious loss and distortion of face details in some output results.At present,to solve this problem,most of the segmentation image's foreground and background,or use the built-in mask generator structure to shield the influence of background on the migration effect.However,there is not much research on the impact details.In order to explore the influence of background on the prospect of gender transfer in Cyle GAN,this paper studies the dynamic characteristics of background influence in gender transfer in Cyle GAN,which provides a new way to improve the quality of gender transfer generated images.In this paper,Python is used as the development language to build a platform for Tensorflow 1.13.1 neural network.The image processing part adopts Open CV 3.4.1.The main work of this paper includes the following contents:(1)To explore the quality of the image generated by gender conversion in different backgrounds,we need to expand the sample set of the original character image to obtain the training set of multiple source domain samples of the background complexity of the regularized image.In the experiment,firstly,through the Deeplab V3+ model of semantic segmentation network,the foreground of male and female character sample sets are segmented respectively.By using ROI region selection and replication technology,images with different background attributes are fused with the foreground of segmented characters,and multiple sets of source domain sets with different background complexity are constructed.In order to simplify the difficulty of sample set construction,the male is used in the experiment Sex as the source domain sample,women as the target domain sample.(2)Build a CycleGAN structure of style transfer network.For the training samples with different complexity,the value of cyclic consistent loss coefficient will have an impact on the effect of final model.In the experiment,the relevant literature is fully investigated and the reasonable value of ? is selected to facilitate the subsequent experiment to better generate high-quality images.The idea of control variable is adopted to ensure that each experiment has only background transformation,and migration training is carried out on the premise that the foreground and network parameters are unchanged.After the same number of iterations,the training models of each experiment are obtained.Finally,the same test set is used to test each model.(3)Data analysis and quality assessment of background and test results.The test results of different models will have different degrees of distortion.In order to choose an image quality evaluation method that can well satisfy the high correlation with subjective scores on different distortion big data sets,the experiment analyzes the image from three aspects of brightness,texture and contrast.Finally,combined with the above three aspects,we use ms-ssim image quality evaluation method to analyze the background and test results,and find the inverse scale characteristics of the background ms-ssim values in the source domain and the target domain.That is to say,the more similar the background of domain X and domain y is,the worse the quality of the image generated by the training model is.A new way to improve the quality of foreground style transfer image is put forward.