Research Of Image Synthesis Method Based On Improved U-Net

Image synthesis is a process of generating photographic images using some form of image description.A deep learning-based image synthesis method uses an existing data set to train a deep network to learn a method of synthesizing a corresponding image.U-Net is a commonly used convolutional neural network structure in deep learning,mainly used for medical image segmentation.In this paper,U-Net is applied to the field of image synthesis,and the defects of U-Net in image synthesis application are improved accordingly.The research content of this paper is as follows:Firstly,in order to increase the parameter capacity of the network,this paper introduces the dense connection structure in DenseNet and the residual structure in ResNet into U-Net,and a dense residual module is proposed.By introducing dense connections and residual structures,U-Net's network parameter capacity has been greatly improved,and the utilization of feature maps in the network has been improved.The network can learn more features and generate more detailed images while maintaining Network training and prediction efficiency.Secondly,in order to improve the quality of the synthetic image of the convolutional neural network,this paper replaces the transposed convolution used in U-Net with Resize-Conv.U-Net uses transposition convolution to enhance the feature map,but the transposition convolution affects the quality of the composite image,resulting in different degrees of checkerboard artifacts in the composite image.The Resize-Conv used in this paper eliminates the checkerboard artifacts present in the composite image to a certain extent,improving the quality of the composite image.Then this paper combines the perceptual loss and Smooth L1 loss function to train the network through supervised learning.Using perceptual loss can make the network learn abstract features of objects,while Smooth L1 loss can restrict the details of synthetic objects,such as color,etc.In the end,the paper carried out two kinds of experiments from semantic images to realistic images and sketch images to realistic images.Experiments on the three datasets of Cityscapes,ZuBuD and CelebA show that in the complex image synthesis task,this paper proposes Optimized networks(IUNs)have obvious advantages in efficiency compared with CRNs,and the quality of synthesized images is slightly better than CRNs.The image is richer in detail than the image synthesized by U-Net,and the object in the synthesized image is more complete.In simple synthesis tasks,IUNs behaves similarly to U-Net,but the images synthesized by IUNs are of higher quality,and there are fewer artifacts in the image.In summary,the improved network proposed in this paper achieves a good balance between synthetic efficiency and synthetic image quality.