| Texture synthesis is one of the fundamental research topics in the field of computer vision,and with the development of multimedia technology,its application areas have been further expanded,such as game modeling,movie rendering,and virtual drawing.Traditional methods have made significant progress in achieving this goal,however,for complex texture patterns,existing methods are not capable of extracting enough features,resulting in incorrect synthesis results.In recent years,deep learning has become a popular method for image processing,in addition to many extended applications.However,neither the traditional texture synthesis methods nor the recently emerged convolutional network(CNN)and generative adversarial network(GAN)approaches for texture synthesis can deal well with the problems of unreality,lack of detail,and distortion of texture structure.Based on this,thesis conducts research around the following aspects.1.Non-homogeneous extended synthesis algorithm based on generative adversarial graphical convolutional networks is proposed.Although the generative adversarial network has good image generation capability,texture feature extraction is not sufficient to capture the global features of the image,so a new texture synthesis method combining adversarial generative network and graph convolutional network(GCN)is proposed.This is achieved by constructing a graph structure in the sample texture image and the synthetic image.Since the texture properties of the texture image are closely related to the intrinsic connection between each pixel value,the graph structure is constructed based on the relationship between pixel values,and the graph node features are extracted by graph convolution so as to obtain the deep dependencies between each pixel value.Sample and synthetic images should have similar graph features.Based on this,thesis designed graph similarity loss metric to measure the intrinsic graph feature differences between sample images and synthetic images,and adopts an end-to-end approach to integrate graph networks into a network framework and combine them with generative adversarial networks to solve the problem that existing methods have difficulty in capturing the global structure of Non-homogeneous textures.2.In order to be able to realize the optimization work of synthesis of various types of textures within a unified framework,especially the optimal synthesis of large scale textures and the synthesis of regular textures,thesis proposes a residual attention module and introduces a generator network to optimize the generative capability network of the network.The attention mechanism can enhance the feature extraction ability of the network for key information and key texture structures of images,so as to improve the detail of texture images synthesis capability,and help the model master the large scale texture structure and reduce the structure distortion problem of large scale texture.3.propose a homogeneous texture synthesis algorithm based on a multi-head mutual self-attention mechanism.In order to reduce over-computation,a synthesis algorithm for homogeneous textures is proposed.which effectively solves the quality problems such as low resolution and insufficient details in texture synthesis.The main idea of the method is to design a multi-head mutual self-attention mechanism,which can be used in the optimized generative adversarial network to accomplish texture synthesis tasks.Unlike self-attention,multi-head mutual self-attention models the inter-influence relationship of each location in feature space and can generate details using cues from all feature locations.Therefore,embedding the multi-head mutual self-attention mechanism into the generator helps the generator to improve the ability to extract detail features and edge features4.The experimental results show that the non-uniform texture extension synthesis method based on generative adversarial graph convolutional networks proposed in thesis can effectively solve the problems of structural distortion and misshapen patterns in the extension synthesis of non-uniform textures.At the same time,it is able to be applied to image texture migration,which synthesizes a different style of the image by migrating the texture of a textured image to another content image.In addition,the texture synthesis method based on the multi-head mutual self-attention mechanism is able to retain the detailed information of the texture intact and the synthesis results are more realistic. |
