Arbitrary Image Texture Synthesis Based On Feature Map Matching Of Convolutional Neural Networks

In nature and real life,texture images can be seen everywhere.Researchers have conducted in-depth research on the use of computer-generated texture images and propose many texture synthesis algorithms.In this thesis,traditional texture synthesis algorithm and texture synthesis algorithm based on convolutional neural network are studied and compared.The shortcomings of existing methods are found: traditional texture synthesis algorithm is not universal for some texture types.Sexuality,the generated texture image may be blurred.Since the input image size of existing method with convolution neural network is fixed and the network structure is fixed,image needs to be restored to original size or image is enlarged using Up Sampling to generate an integer multiple.Therefore,the motivation of this thesis is to use a convolutional neural network to synthesize multi-scale texture images from a given sample.In this thesis,a texture synthesis network framework based on convolutional neural network is designed.This framework consists of two parts: the texture generation network and the loss function calculation network.The texture generation network model constructed is based on Generative Adversarial Networks(GAN),which includes generators and discriminators.The result of the texture generation network can be extended from the inside to the periphery,expanding the scale of the generated texture image to any multiple size.The generator uses a neural network model with dilation layers and the discriminator uses the Patch GAN network model.In the loss function calculation network,the VGG-19 is taken as the main model.Swap algorithm is designed to optimize the loss function.Swap algorithm can perform operations on the texture image feature map,and match the texture features between the original image and the generated image.Feature map after the nearest neighbor matching is applied to the calculation of the loss function.This method was tested on Describing Textures in the Wild(DTD)and Places365 datasets.Compared with other existing convolutional neural network-based texture synthesis methods,the proposed method outperforms other methods in generating multi-scale texture images and it is the effectiveness methods.