基于深度学习的自适应水墨画计算美学评估

Self-Adaptive Computational Aesthetic Evaluation of Chinese Ink Paintings Based on Deep Learning

近几年艺术作品的计算美学评估已成为一个热门的研究方向.但现有工作主要研究照片和油画,关于水墨画的定量审美评估却鲜有尝试.水墨画通过水墨相调和笔法变化来表现画面,因而在视觉特征、语义特征和审美准则上与照片和油画有显著不同.针对此问题,采用深度学习技术,提出一种自适应的水墨画计算美学评估框架.该框架首先构建水墨画图像美学评价基准数据集;然后根据水墨画审美标准提取全局与局部图像块作为多路输入,并设计一种多视角并行深度卷积神经网络来提取深度审美特征;最后基于水墨画的题材查询机制,构建自适应深度审美评估模型.实验结果表明,文中包含6个并行题材卷积组的多视角网络架构相较基础VGG16架构有较高的审美评估性能,提取的深度审美特征明显优于传统手工设计特征,其自适应模型评估结果与人工审美评价之间达到0.823的皮尔森高度显著相关,且均方误差为0.161.此外,干扰实验表明,文中的网络对构图、墨色和纹理3个绘画要素较为敏感.该研究将不仅为国画计算美学评估提供了一个基于深度学习的参考框架,而且有助于进一步探索人类审美感知与水墨画中深度学习特征之间的关系.

Computational aesthetic evaluation of artworks has become an active research direction in recent years.However,current works mainly focus on oil paintings and photographs,there have been few attempts in quantitative aesthetic evaluation of Chinese ink paintings.Chinese ink painting uses ink blended with water and a variation of brushwork to depict picture,which differs significantly from photographs and oil paintings in visual features,semantic features,and aesthetic principles.Aiming at this problem,we propose a framework of self-adaptive computational aesthetic evaluation of Chinese ink paintings based on deep learning technique.Firstly,we build an aesthetic evaluation standard dataset for ink painting images.Sec-ondly,according to aesthetic principles of Chinese ink paintings,we design a multi-view parallel deep neural network by taking global images and local patches as multi-column inputs to extract deep aesthetic features.Finally,we build a self-adaptive deep aesthetic model of Chinese ink paintings based on subject query mechanism.Experimental results show that,compared with the basic VGG16 architecture,our multi-view network that contains six paralleled subject convolution groups has higher aesthetic evaluation performance,the deep aesthetic features outperform the traditional hand-crafted features,and our proposed self-adaptive model can predict human aesthetic decision with highly significant Pearson correlation of 0.823,with mean squared error of 0.161.Moreover,interference experiments show that our network is sensitive to painting factors including composition,ink color,and texture.Our work not only offers a deeply-learned-based ref-erence framework for quantitative aesthetic evaluation of Chinese paintings,but also reveals the relationship between human aesthetic perceptions and deeply-learned features extracted from Chinese ink paintings.