可计算的图像美学分类与评估  

Computational Aesthetics of Image Classification and Evaluation

作　　者：王伟凝[1] 蚁静缄[1] 徐向民[1] 王励[1] 

机构地区：[1]华南理工大学电子与信息学院,广州510640

出　　处：《计算机辅助设计与图形学学报》2014年第7期1075-1083,共9页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61171142);广东省战略性新兴产业核心技术攻关项目(2011A010801005);中央高校基本科研业务费(2014ZZ0036)

摘　　要：可计算图像美学的研究目的是希望计算机能够模拟人类的视觉系统与审美思维,对图像进行美学价值的判断.以摄影图像为研究对象,首先通过显著区域检测将图像划分为整体区域和关键区域,提取有效的低层视觉特征、高层美学特征和区域特征;然后以人对图像的主观美感评分为基础,采用机器学习的方法建立图像美感等级分类器和美学分数评估模型,实现了高、低美感图像的自动分类,并能对图像的美学分数进行自动评估.在美学图像数据库上的实验测试结果显示,高、低美感分类模型的平均分类准确率为75.37%;美学分数自动评估模型的结果与人的主观审美评分之间的相关性为0.790,均方根误差为0.244;该结果符合人类对图像的审美感知.同时,通过与其他算法进行对比实验,结果表明文中算法更为有效.Assessing image aesthetic appeal referring to principles of the nature, visual appreciation and psychology of human beings, is a highly subjective task. Computational image aesthetics aims to assess the aesthetics of images automatically. In this paper, we design a comprehensive method for aesthetic evaluation focusing on digital images. First, key regions of images are extracted through salient region detection method. Then a set of discriminative low-level visual features and high-level aesthetic features from the global images are extracted, as well as regional features that characterize the key regions. Based on the extracted features and human aesthetics ratings, the image aesthetic classifier and the aesthetic score prediction model are built through machine learning methods. As a result, the aesthetic category （highly aesthetic VS low aesthetic）, as well as the aesthetic score, of an image can be evaluated automatically through our models. The experiment results demonstrate that our image aesthetic classifier achieves a promising classification accuracy of 75.37%. In the mean time, our aesthetic score prediction model got a correlation coefficient of 0.790 and RMSE of 0.244 on images’ aesthetic score between automated assessment and human beings’ subjective aesthetics evaluation. The results show that the method proposed in the paper is competitive with previous approaches.

关 键 词：图像美学 美感分类 美学分数评估 美学特征提取 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]