Research On Image Synthesis Method Based On Improved U-net

Image synthesis is an important research direction in deep learning and visual computing,and it has wide application prospects and value in today's social life.It generally extracts and analyzes the features of the input original image,then makes specific constraints and transformations on these features to obtain new features,and finally uses the new features to synthesize and output new images.Traditional image synthesis methods are generally based on artificially set features such as color histograms,texture information,shape information,and gray information,and then combine prior knowledge to conduct image synthesis research.This type of method is too dependent on artificially set shallow features and ignores the deep features and spatially related features of the image.Therefore,it is difficult to synthesize a satisfactory new image,so that it cannot be applied in the actual environment.In recent years,image synthesis methods with deep learning technology as the core have achieved great results in many industries.This type of method mainly uses deep neural networks to build corresponding image generation models for new image synthesis.U-net and generative adversarial networks are two deep network models commonly used for image synthesis.However,U-net still has certain limitations.U-net is a U-shaped network with bilateral symmetry,the number of convolutional layers in the network will increase exponentially while increasing the output image resolution,which will lead to deeper network layers,make the training of the network more difficult,degrade the learning performance of the model,and In the end,the resulting image has a lower quality.Therefore,it is very necessary and meaningful to improve and optimize the structure of U-net.In order to solve the above problems,this paper proposes two improved methods for U-net structure.The first is the U-net model based on high-speed residual blocks for real scene image synthesis.This method improves the image synthesis performance of U-net by embedding high-speed residual blocks in the stack convolution layer in the vertical direction of U-net.The second is the U-net model based on swish threshold residual block,which is used for the automatic coloring of black-and-white sketches.The method embeds swish threshold residual block in the vertical stacking convolution layer and horizontal skip connection of U-net to improve the image synthesis performance of U-net.In order to prove the superiority of this method,the above two improved methods have been verified by a large number of experiments on the publicly standard datasets respectively.The experimental results show that the methods proposed in this paper have better performance on the corresponding image synthesis tasks.Finally,this paper further explores the above second method,and uses the U-net based on the swish threshold residual block as the generator in the generative adversarial networks for the automatic coloring of black-and-white sketches The experimental results show that the method has excellent automatic coloring performance.