| Since the appearance of texture synthesis technology,focusing on the reconstruction and generation of various natural textures,the goal is to enlarge the image size,make the generated results clear and rich in details.Texture synthesis not only plays a positive role in promoting the development of virtual surgery,VR and multimedia technology,but also can be used as the basis of other downstream tasks of computer vision.Therefore,texture synthesis has important practical and theoretical significance.Deep learning has been widely used in texture synthesis,but the generated results of images with strong structure are flawed due to the strong position relationship between textures.In order to the solve texture synthesis tasks,it is so important to study structural texture and propose a generation model suitable for various texture types.Because of the powerful feature extraction and reconstruction capability of generative adversarial network,a texture algorithm based on generative adversarial networks is proposed in this thesis,which solves the problem of enlarging image size and preserving image details better.The sub-pixel convolution operation and multi-scale channel attention module were used to improve the performance of the generator.The optimal transmission distance of features between images was calculated by convolutional neural network,and the pixel loss between images and samples was calculated to further constrain the statistical data between images.The main work of this thesis is as follows:1.In this thesis,a data preprocessing method is set for the input data and training set for the task of generating a large-size result graph from the input noise image and retaining the detail information of the texture sample graph.Based on the generative adversarial networks,this thesis proposes a multi-scale texture synthesis model.Channel attention network in learning the distribution of the sample data,at the same time,through the sub-pixel convolution layer to expand the size of the input image noise,as a result,multi-scale channels with attention weighting attention modules are characteristic figures,and this can further enhance the performance of the generator.Experimental results show that this method avoids the problems of detail loss and local distortion,and can obtain high quality generated images.2.Based on the problems of fuzzy details and overall structure deviation of some sample images in multi-scale channel attentional texture synthesis model,this thesis proposes a texture synthesis model combined with optimal transmission distance to solve the problem of generating sample structure deviation.Sliced wasserstein distance feature space can be transformed to the depth,and also can capture more complete distribution.So this thesis uses VGG19 network to extract the feature map information of images,and further restrains the correlation of statistical features between two images by calculating the sliced wasserstein distance between images and texture samples.Therefore,a loss model based on the optimal transmission distance is proposed in this thesis,which includes the loss of generative adversarial,pixel loss and sliced wasserstein loss.The results show that the loss model achieves better visual effects,and the generated image achieves better quality in details and overall structures.Through experimental verification,the method proposed in this thesis can expand the image size and generate sample images of any size,avoid structural deviation and detail loss,obtain better visual effects,and further increase the interpretability of the network. |
