Research On Unsupervised Image-to-Image Translatio

Image to image translation(I2I)is a hot topic in academic circles in recent years.Many images processing,computer graphics and computer vision problems can be regarded as image to image translation tasks.This translation requires learning to map one visual representation of a given input to another.Image to image translation using the Generation Adversarial Network(GAN)has been deeply studied and applied to various tasks,as well as to image synthesis,super resolution,virtual fitting,virtual live broadcast,and other tasks in the actual industry.The traditional I2 I translation can achieve extraordinary results between two more similar datasets,but the quality of the results may be low when the shapes and edges of the source image domain and the target image domain differ greatly.In addition,when the training model does not use paired data,the translation from the source domain to the target domain may be unstable during the training process.Therefore,this paper deeply analyzes how to discard the shapes and edges of the source image,preserve the layout and color of the image,and preserve the quality of the generated image in the I2 I task when the shapes and edges of the source image domain and the target image domain differ greatly.Based on this,two long-domain unsupervised I2 I translation algorithms are proposed.The specific research contents are as follows:(1)Since most of the existing I2 I translation algorithms are only applicable to scenes where image edge details need to be retained in translation tasks,and the image quality generated when the shapes and edges of the source image domain and the target image domain differ greatly,we propose a new unsupervised image to image translation method.This method includes a new skip path search block and a new loss function to maintain image similarity.The skip path search block helps the model expand the search scope in the target domain and guides the feature maps to finally migrate to the corresponding image closest to the optimization target.The proposed new Sketch Loss function is more suitable for maintaining the human visual similarity of images in the migration with long distance between domains.In addition,to preserve more detailed source domain images in the generated images in long-domain translation,we propose a spatial modulation method to transfer clearer details.The experimental results show that,compared with the existing advanced models with fixed network structure and super parameters,this method has advantages in image quality,semantic similarity between the input image and the generated image,and the generated image inherits the structure of the input image in terms of color and shape details.(2)When the shapes and edges of the source image domain and the target image domain differ greatly,the existing I2 I translation algorithms are difficult to simultaneously maintain the shape of details,image structure layout,and image global color in the translation task.To solve this problem,we propose a new unsupervised I2 I translation method,which includes a new path search modulation block and a loss function to limit the dependence of the generated image on the edge of the input image.The path search modulation block helps us to control the transfer of local shape details according to the local mapping of the input image during the generation of the target image,so that the model can better discard the edges and textures that should only appear in the source domain but not in the target domain during the generation process.In addition,the path search modulation block also well inherits the capture capability of the skip path search block for image color layout.The new loss function limits the transfer range of local mapping,so that the target domain can generate high similarity images depending on a small amount of local mapping of the source domain.The experimental results show that the method has advantages in the similarity of source image and target image,the quality of target image,and the edge of the generated image is more reasonable in human visual effect.