基于卷积神经网络的水墨画合成方法  

Vexture synthesis of ink painting based on convolutional neural network

作　　者：彭玉元[1,2] 赵朋朋 武有能 谢兵兵[2] PENG Yu-yuan;ZHAO Peng-peng;WU You-neng;XIE Bing-bing(School of Art and Design,Guilin University of Electronic Technology,Guilin 541004,China;School of Computer Science and Information Security,Guilin University of Electronic Technology,Guilin 541004,China)

机构地区：[1]桂林电子科技大学艺术与设计学院,广西桂林541004 [2]桂林电子科技大学计算机与信息安全学院,广西桂林541004

出　　处：《桂林理工大学学报》2019年第1期204-209,共6页Journal of Guilin University of Technology

基　　金：国家自然科学基金项目(61662018; 61661015);广西自然科学基金项目(2016GXNSFAA380153; 2015GXNSFAA 139294);广西云计算与大数据协同创新中心项目(YDQ17001)

摘　　要：针对传统基于图像的水墨画绘制方法生成的图像只具备水墨画的一些基本特征,不能给其指定某一种风格,使得生成的图像显得呆板、缺乏意境层次的表达的情况,提出一种基于卷积神经网络的水墨画合成方法。该算法调整一个训练完成的卷积神经网络模型的结构,定义图像在卷积神经网络模型中的卷积层映射的特征响应的表示,以及特征响应之间的相互关系表示。先对照片作对比度增强预处理,然后在一张随机的图像上,匹配照片的特征响应来获取内容信息,匹配水墨画的特征响应相互关系来获取风格信息。实验表明,基于卷积神经网络的水墨画合成方法可以生成效果较好的水墨画图像,既保留了原照片轮廓信息,又带有原水墨画整体纹理信息,对水墨画灰度图像的风格合成效果出色。In the past, image by image-based rendering method in ink paintings only displays some basic characteristics, and can not specify a particular style. So the generated image looks a little dull and has no artistic conception or impression. In this situation, a new ink painting synthetic method based on convolution neural network is proposed. An adjustment is made on the structure of a trained convolutional neural network, defining a representation of feature response of image convolution layer mapping which is in the convolutional neural network model, and the representation of the correlation in the feature responses. The contrast enhancement preprocessing for the image shall be done on the first. Then, on a random image, the content information is obtained through matching the feature response, and the style information obtained through matching the feature response correlation of the ink painting. Experimental results show that ink synthetic method based on convolution neural network can generate better ink performance image, with not only outline information of the original photo, but also overall texture information of the original ink painting. A novel method for ink painting style to a photo is proposed. The method for gray image synthesizing in ink paintings has good performance.

关 键 词：水墨画 卷积神经网络 纹理合成 图像增强 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]