基于创意序列数据的艺术风格学习与绘制方法研究

Artistic stylization in non-photorealistic rendering enables users to stylize pictures with the appearance of traditional art forms, such as pointillism painting, line sketching, or brush stroke drawing. Among them, the brush stroke drawing is one of the widely used art styles across different cultures in history. In computer-generated painterly rendering, the stroke placement is a big challenge and significant efforts have been made to investigate how to draw a stroke with realistic brush texture in a desired shape and how to organize multiple strokes. ..The most straightforward approach for painterly rendering would be physics-based painting, i.e., giving users an intuitive feeling just like drawing with a real brush. These virtual brushes can be used to draw various styles of strokes with a digital pen or mouse. However, it is very complex to control a virtual brush. Furthermore, since the computational cost is often very high to achieve satisfactory visual effects to human eyes, some physics-based painting approaches rely on graphics processing units (GPUs) for obtaining reasonable performance...In this proposal, we believe that the human behavior for artistic creativity contents more information on the artistic stylization. Our proposed system is built on the reinforcement learning (RL) framework to dynamic stroke-based stylization. Our main contributions includes that (1) we develop a creative process motion capture system to capture large scale complex experts' samples, (2) we learn the state features through deep networks, (3) we learn the reward function from artists' samples towards IRL, (4) we propose the HGS-PGPE method to learn agent's policy. Finally, to demonstrate through experiments the effectiveness of our proposed approaches, we implement a web application for agent's automatic stylisation of converting photographs into stroke drawings.

基于学习的智能图像风格绘制是当前计算机图形学与多媒体特别是艺术风格化领域的一个热门研究课题。图像风格学习算法主要研究利用计算机图形学与多媒体及机器学习方法对真实样本数据进行自动艺术绘制智能处理。目前主流方法主要针对艺术图像的静态样本进行学习。然而，由于静态数据类中包含的信息是平面化、局部化和非连续化的，难以保证风格化处理的全局一致性。本项目旨在针对创意过程中的大规模复杂数据，从序列任务学习理论上提出一套智能艺术风格绘制的理论模型、设计方法以及优化方法三个层面。期待完成以下研究工作：（1）行为序列数据的精确捕捉，状态分析与特征提取；（2）基于逆向强化学习的艺术特征结构化组织与反馈式评价函数构建；（3）基于参数探索的风格化动态策略方法。最终实现面向移动互联网的自动艺术风格绘制辅助系统。本课题的研究不仅具有理论上的创新而且还具有实际上的巨大应用和价值。

基于学习的智能图像风格绘制是当前计算机图形学与多媒体特别是艺术风格化领域的一个热门研究课题。图像风格学习算法主要研究利用计算机图形学与多媒体及机器学习方法对真实样本数据进行自动艺术绘制智能处理。本课题已完成的主要内容包括，针对创意过程中大规模复杂序列数据，从解决序列任务的角度构建完成了一套艺术风格化动态决策的理论模型，具体包括状态设计、评价函数构建及策略探索方法三个层面，并通过理论证明、模拟系统实现与评价进行验证。在项目期间，提出了一系列具体算法：提出创意过程中真实动作序列数据与画作生成真实序列数据的采集方法。用于匹配及其大规模环境中高维状态变量的自动表达方法。为策略学习系统构建一套摆脱人工设计状态特征的方法。为了评价风格评价的一致性，提出了艺术风格行为的特征的层次化组织及反馈式评价度量方法。为了保障模拟创意行为的有效性，探索针对复杂长序列任务的策略模型，使之即使在真实专家数据有限的情况下，也能够有效避免策略梯度估计方差大的问题，保证风格化策略学习性能。综上所述，我们取得了一定的成果，并发表了高水平的学术论文。