摘要
Abstract
The state-of-the-art analyzed texture synthesis techniques are divided into non-parametric and parametric methods, which contribute to the current corresponding research on computer vision. By summarizing and comparing their principles,structures,develop-ment trends,advantages and disadvantages,a non-parametric method based on graph-cut model and a parametric method based on P&S model are analyzed in detail. In addition,the structures and principles of Convolution Neural Network ( CNN) based on deep-learning which are widely applied in image-process filed are also discussed. Finally,a new texture synthesis model VGG-19 is introduced,which is the combination of CNN-based Caffe network with VGG model that obtained high scores in the 2014 ImageNet classification and ob-ject detection competence. The VGG-19 model can be also used to analyze human visual process. The analyzed results show the facts as below. Non-parametric methods can synthesize high-quality textures of various kinds with high speed. Parametric methods are appropri-ate for being used as analysis tools. CNN applied in parametric methods can greatly reduce the time period of designing and adjusting fea-ture representations and parameters and improve the synthesized results synchronously,which is proved to be valuable tools for analyzing theory and realizing applications on texture-synthesis work.关键词
分析式纹理合成法/非参数法纹理生成/参数法纹理生成/深度学习/卷积神经网络/VGG-19Key words
analyzed texture synthesis method/non-parametric texture generation/parametric texture generation/deep learning/convolu-tional neural network/VGG-19分类
信息技术与安全科学
