分析式纹理合成技术及其在深度学习的应用  

Analyzed Texture-synthesis Techniques and Their Applications in Deep Learning

作　　者：李宏林 

机构地区：[1]日本山梨大学大学院生命情报系统系

出　　处：《计算机技术与发展》2017年第11期7-13,共7页Computer Technology and Development

摘　　要：当前国际主流的非参数和参数法分析式纹理生成技术,对于计算机视觉领域的图像纹理合成具有一定的借鉴意义。在概括总结与比较分析式纹理生成技术原理、框架结构、应用发展趋势及其优缺点的基础上,分析了基于graph cut模型的非参数法、基于P&S模型的参数法两种典型的纹理生成技术以及广泛应用于图像处理领域的深度学习新技术—卷积神经网络(CNN)的结构与原理,进一步讨论了以基于CNN的Caffe网络框架及在2014年Imag Net图像分类和目标识别大赛上取得优异成绩的VGG模型为基础的分析式纹理生成模型VGG-19的工作原理及其在人脑视觉分析研究方面的应用。分析结果表明:相对于普通参数法和基于CNN网络模型的参数法,非参数法具有更快的处理速度,可生成更高视觉质量与更多种类的目标纹理图;参数法适合作为纹理合成领域的分析研究工具;卷积神经网络应用到参数法中,可大幅缩短特征量设计与参数调整周期并提高合成效果,进一步提升了参数法作为理论分析和应用实现工具的价值。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 ate 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-19 

分 类 号：TP37[自动化与计算机技术—计算机系统结构]