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.
引文
[1]Strassmann S.Hairy brushes[J].AcmSiggraph Computer Graphics,1986,20(4):225-232.
[2]余斌,孙济洲,白海飞,等.基于纸的物理建模的水墨画扩散效果仿真[J].系统仿真学报,2005,17(9):2305-2309.
[3]Lee J.Diffusion rendering of black ink paintings using new paper and ink models [J].Computers & Graphics,2001,25(2):295-308.
[4]Chu N S H,Tai C L.Real-time painting with an expressive virtual Chinese brush[J].IEEE Computer Graphics & Applications,2004,24(5):76-85.
[5]宋瑞丽,孙济洲,孙美君.3D水墨线皴渲染的仿真研究与设计[J].微处理机,2008,29(1):82-85.
[6]石永鑫.中国水墨画中扩散与多笔次叠加效果的仿真[D].天津:天津大学,2003.
[7]石永鑫,孙济洲,张海江,等.基于粒子系统的中国水墨画仿真算法[J].计算机辅助设计与图形学学报,2003,15(6):667-672.
[8]张海江,王秀锦,孙济洲,等.应用分形仿真水墨扩散轮廓[J].计算机辅助设计与图形学学报,2004,16(4):555-558.
[9]Yu J H,Luo G M,Peng Q S.Image-based synthesis of Chinese landscape painting[J].Journal of Computing Science and Technology,2003,18(1):22-28.
[10]李丹.中国山水画的计算机绘制[D].天津:天津大学,2004.
[11]曹毅.基于图像的中国水墨画绘制方法的研究[D].长春:吉林大学,2012.
[12]Liang L Y,Jin L W.Image-based rendering for ink painting[C]//IEEE International Conference on Systems,Man,and Cybernetics.2013.DOI:10.1109/smc,2013,674.
[13]金炜炜.非真实成像的水墨画绘制模型若干技术研究[D].杭州:浙江工商大学,2013.
[14]Gatys L A,Ecker A S,Bethge M.A neural algorithm of artistic style[J].arXiv preprint axXiv:1508.6576,2015.
[15]Sermanet P,Eigen D,Zhang X,et al.Overfeat:integrated recognition,localization and detection using convolutional networks [J].arXiv Preprint arXiv:1312.6229.2013.
[16]Long J,Shelhamer E,Darrell T.Fully convolutional networks for semantic segmentation[J].Computer Science,2014,79(10):1337-1342.
[17]刘丽,匡纲要.图像纹理特征提取方法综述[J].中国图象图形学报,2009,14(4):622-635.
[18]Gatys L A,Ecker A S,Bethge M.Texture synthesis using convolutional neural networks [C]//Advances in Neural Information Processing System,2015:262-270.
[19]何晨阳,周孟然,闫鹏程,等.图像处理与BP神经网络在矿井火灾隐患识别中的应用[J].桂林理工大学学报,2016,36(3):615-618.
