基于CGAN的中国山水画布局可调的仿真生成方法

以往的山水画计算机仿真由于未从山水画整体布局的角度进行研究,难以实现完整的画作生成.针对上述问题,文中提出布局引导、可实现完整画作生成的中国山水画仿真方法.基于山水画的绘制特点设计可行的布局标签图结构,用于表达山水画的构图形态和要素.借鉴条件生成对抗网络(CGAN)的思想,针对山水画的布局和笔触特点,设计并训练多尺度特征融合的网络结构(MSFF-CGAN),实现布局标签图到仿真山水画这一异质生成过程.同时针对网络训练过程中布局标签图数据稀缺的问题,采用语义关联的颜色像素聚类算法快速生成标签图.为了提高生成图的艺术真实感,引入MemNet超分辨网络增强生成图的纹理细节.实验表明,文中方法生成的仿真山水画具有较好的完整性和艺术真实感,不仅可以应对简单的手绘涂鸦式草图,还可以通过在布局空间的编辑操作,达到对画作空间进行编辑的效果.

Layout Adjustable Simulated Generation Method for Chinese Landscape Paintings Based on CGAN

Creating a complete landscape painting via computer simulation is difficult without studying from global layout viewpoint. To address this issue, a layout-guided Chinese landscape painting simulation method for a complete painting generation is proposed. The characteristics of landscape paintings are taken into account in the design of feasible structures of layout label maps. Composition forms and elements of landscape paintings can be depicted using those structures. On the basis of condition generative adversarial network (CGAN) approach, a multi-scale feature fusion CGAN (MSFF-CGAN) is designed based on layouts and touches of landscape paintings. The proposed network is trained to accomplish heterogeneous transfer from a layout label map to a simulated landscape painting. To deal with rare availability of layout label maps for network training, a color pixel clustering algorithm with semantic correlation is used. In order to enhance the artistic reality of the generated landscape painting, a super resolution network named MemNet is incorporated to refine the texture details. Experimental results show that the proposed method is superior to existing methods in both integrity and artistic reality. Moreover, the proposed method can be used to handle simple graffiti sketches and modify simulated landscape paintings by editing label maps.