局部色彩可控的中国山水画仿真生成方法  

Local Color Controllable Chinese Landscape Painting Simulation Generation Method

作　　者：林锦[1] 陈昭炯[1] 叶东毅[1] LIN Jin;CHEN Zhao-jiong;YE Dong-yi(College of Mathematics and Computer Science,Fuzhou University,Fuzhou 350108,China)

机构地区：[1]福州大学数学与计算机科学学院,福州350108

出　　处：《小型微型计算机系统》2021年第9期1985-1991,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61672158)资助;福建省自然科学基金项目(2018J01798)资助。

摘　　要：本文针对布局标签图到仿真中国山水画的转换生成问题进行研究,现有的基于条件生成对抗网络(CGAN)的方法存在色彩和语义失真、网络结构参数量较大的问题.针对这些问题,提出一种局部色彩可控的中国山水画仿真生成方法.方法首先提出并设计了一种面向山水画的多语义标签图作为交互方式,根据内容、技法、颜色3个语义层次归纳出山水画中对象的类别,相应地设计了面向手绘山水画原作的多语义标签图的分层分割生成算法,用于构造"手绘山水画-多语义标签图"数据对以作网络训练之用;其次,提出了轻量化的多尺度颜色类别关注的条件生成对抗网络MS3C-CGAN,引入空间自适应归一化残差块、双线性上采样结构简化并重构原有的UC-Net生成器,将生成器的参数量减少了24.45%.对比实验结果表明,本文方法仿真生成的中国山水画更具色彩艺术真实感、语义内容更为准确,同时通过编辑布局标签图可控制生成山水画中植被的色彩,可应用于艺术教育、设计模拟等领域.This paper studies the problem of translation from layout label maps to simulated Chinese landscape paintings.The existing methods based on conditional generation adversarial networks have problems of color and semantic distortion,large amount of network structure parameters.Aiming at these problems,a local color controllable Chinese landscape painting simulation generation method is proposed.The method first proposes and designs a multi-semantic label map for landscape painting,which is used as an interactive method.According to the three semantic levels of content,technique and color,the categories of objects in the landscape painting are summarized.The hierarchical segmentation generation algorithm of multi-semantic label maps for original hand-drawn landscape paintings is designed to construct the"hand-drawn landscape painting-multi-semantic label map"data pair for network training;Secondly,a lightweight Multi-Scale Color Class Concerned Conditional Generation Adversarial Network(MS3C-CGAN)is proposed.The SPatially-Adaptive(DE)normalization residual block,bilinear upsampling structure is introduced to simplify and reconstruct the original UC-Net generator,which reduces the amount of generator’s parameter by 24.45%.The comparative experiment results show that the Chinese landscape paintings generated by the simulation method in this paper have more colorful artistic reality,and more accurate semantic content.At the same time,the color of the vegetation in the landscape paintings can be controlled by editing the layout label map,which can be used in art education/design simulation and other fields.

关 键 词：中国山水画仿真 布局标签图 局部色彩可控 CGAN 空间自适应布局归一化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]