基于三支特征表示的抽象画情感聚类分析

Affective clustering for abstract paintings with three-way features

针对绘画图像情感标注所需资源巨大的问题,设计一种针对抽象画图像的情感聚类方法。提出一种基于三支决策的颜色特征表示方法和纹理特征表示方法,结合改进的深度学习模型,从抽象画图像中提取颜色特征、纹理特征和高层语义特征;使用多核k均值算法,自适应地融合3种特征,实现图像的情感聚类分析。实验结果表明,在MART和Deviant-Art数据集上,与4种基准方法相比,提出方法在准确度、Fowlkes-Mallows指数和标准化互信息上分别平均提高了30、23和49个百分点。提出方法在抽象画图像的情感聚类分析应用中表现出色,这也为其它绘画作品的无监督情感分析研究提供了参考。

To address the problem that the sentiment annotation of paintings requires a significant cost,a sentiment clustering method for abstract paintings was designed.A color feature representation and a texture feature representation,both based on three-way decisions,were proposed along with an enhanced deep learning model.These representations were utilized for extracting color,texture,and high-level semantic features from abstract painting images.Subsequently,the three features were adaptively fused using a multi-kernel k-means algorithm,resulting in sentiment clustering outcomes for the images.Experimental results show that compared with four benchmark methods on the MART and DeviantArt datasets,this method improves the accuracy,Fowlkes-Mallows index,and normalized mutual information by an average of 30,23,and 49 percentage points,respectively.The method performs well in the application of sentiment clustering analysis of abstract paintings,which also provides a benchmark for unsupervised sentiment analysis studies of other art paintings.