毕业生知识库

绘画作为一种重要的艺术表现形式，承担了审美认知、审美教育、审美娱乐等多种社会功能。近年来随着人工智能的不断发展，深度学习技术深刻地影响了艺术创作领域。绘画图像生成是一类通过参考真实绘画风格进行新绘画内容创作的技术，在社交、娱乐与辅助艺术创作等方面具有广泛的应用价值。现有的绘画图像生成方法普遍忽视了笔触在绘画中的重要影响，另外在生成高质量绘画样本的同时样本多样性较差。针对机器进行艺术创作时遇到的困难，有研究者提出平行艺术框架。平行艺术旨在构建艺术创作中机器与人的伙伴关系，针对艺术创作过程建立平行系统，为解决机器进行绘画生成所面临的上述问题提供了新思路。

本文以平行艺术为指导，对绘画图像生成任务中样本多样性、绘画重建与风格迁移三个方面展开研究，提出了基于扩散模型的绘画图像生成方法，基于笔触渲染的绘画重建方法与基于笔触特征统计的风格对齐模块，尝试解决生成样本多样性差，绘画重建笔触纹理不真实与风格迁移中笔触风格缺失三个问题。本文的主要工作如下：

• 提出了一种基于扩散模型的绘画图像生成方法。以画家风格作为特有属性，实现了从噪声开始通过迭代去噪过程生成具有特定画家风格的绘画图像。对比实验结果表明，该方法提高了绘画生成样本的多样性。

• 提出了一种基于笔触渲染的绘画重建方法。建立笔触的参数化模型，通过可微渲染器构造参数化笔触到像素域的映射，以更新独立笔触参数的形式进行在线图像优化，实现了绘画图像的笔触重建。

• 提出了一种基于笔触特征统计的风格对齐模块。基于笔触特征的损失函数与像素级风格和内容损失进行联合优化，实现了风格化过程中笔触特征的精确分离和自由控制；同时基于笔触特征对齐, 以名画的笔触为参考，实现了绘画笔触风格迁移。

As an important form of artistic expression, painting undertakes various social functions such as aesthetic cognition, aesthetic education, and aesthetic entertainment. In recent years, with the continuous development of artificial intelligence, deep learning technology has profoundly affected the field of artistic creation. Painting image generation is a technology for creating new painting content according to a specified painting style, which has a wide range of application values in social, entertainment, and auxiliary art creation. Existing painting image generation methods generally ignore the important influence of brushstrokes in painting and generate high-quality painting samples with poor sample diversity. In view of the difficulties encountered in machine art creation, some researchers put forward the parallel art framework. Parallel art aims to build a partnership between machines and humans in artistic creation, establish a parallel system for the process of artistic creation, and provide new ideas for solving the above problems faced by machines in painting creation.

Guided by parallel art, this thesis studies sample diversity, painting reconstruction, and style transfer in the task of painting image generation, and proposes a painting image generation method based on the diffusion model, a painting reconstruction method based on brushstroke rendering, and a brushstroke-based painting reconstruction method. Try to solve the three problems: poor diversity of generated samples, unreal brush stroke texture of painting reconstruction, and loss of stroke style in stylization. The main work of this thesis is as follows:

· Propose a method for painting image generation based on the diffusion model. Taking the painter's style as a unique attribute, a painting image with a specific painter's style is generated through an iterative denoising process starting from noise. The comparative analysis of the results with the existing methods shows that this method solves the problem of poor diversity of painting samples.

· Propose a painting reconstruction method based on stroke rendering. The parametric model of brushstrokes is established, the mapping of parametric brushstrokes to the pixel domain is constructed through a differentiable renderer, and online image optimization is carried out in the form of updating independent brushstroke parameters, thereby realizing the brushstroke reconstruction of painting images.

· Propose a style alignment module based on stroke character statistics, which supports joint optimization with pixel-level style loss, and realizes precise separation and free control of stroke characteristics in the stylization process. Based on the method of brushstroke feature alignment, with the brush stroke features as a reference, the style transfer of painting brush strokes is realized.