Research On View Translation Based On Generative Adversarial Network

View translation technology can use a source view image to generate images from multiple target view,which helps save the cost of acquisition equipment and time.Based on the advantages of generative adversarial network（GAN）in the field of image processing,the current view translation algorithms are mostly implemented in the form of GAN.Existing GAN-based models can achieve a small angle of view translation.However,when the view span is large,significant differences appear between the source and target domains,resulting in the degradation of the generated image quality.In addition,these models have a large number of parameters,consume a lot of memory,and cannot be applied to more scenarios.Moreover,some models can only complete the translation between two views.Generating multiple view images requires training multiple models,which results in low translation efficiency.In view of the above problems,this paper proposes three different GAN-based view translation algorithms,including:（1）Aiming at the problem of poor image quality when the view span is large,this paper proposes a two-stage local and global information guided generative adversarial network（LGGAN）.The local information processing module captures the local details of objects by combining multi-scale features and semantic mapping.The global information processing module facilitates the generation of discriminative regions in the image through global pooling.In addition,dilated convolution and skip connections are integrated into the parameter-sharing discriminator to judge real/fake image pairs.Experiments show that the network can complete high-quality image translation between two views with a large span,and surpass algorithms such as Selection GAN in subjective vision and six objective metrics.（2）Aiming at the problem that view translation models have many parameters,this paper proposes a one-stage semantic and edge features guided generative adversarial network（SEGAN）.The proposed generator includes a semantic encoding module,an image generation module and an edge generation module.The semantic encoding module extracts multi-scale semantic features and passes them into the encoding stage of the image generation module.The decoding stage of the image generation module integrates the new attention module and passes the corresponding features into the edge generation module.Finally,the two modules generate target image and target edge image,respectively.Furthermore,a novel multimodal discriminator is designed,which integrates dilated convolution and attention module to judge real/fake image pairs and real/fake edge images,respectively.Experiments show that SEGAN can achieve higher image quality with 33%of the LGGAN parameters.（3）Aiming at the low efficiency of using a two-view image translation model to complete the multi-view image translation task,this paper proposes a multi-view image translation algorithm based on generative adversarial network and Transformer（MVTGANT）.The Transformer structure is introduced in the encoding stage of the generator to facilitate learning the overall content of the image.At the same time,the adaptive instance normalization layer is used to add view labels to guide the generation of target images with specific view attributes.In addition,an auxiliary classification layer is introduced into the discriminator for view prediction,and the view prediction loss is added for training.Experiments show that the network can accurately generate target images from multiple views with a single-view image,and is at the optimal value on the L₁ and SSIM metrics.