Learning From Multi-dimensional Correlation For Visual Art Image Classification

Visual arts involves a variety of art forms including painting,architecture,photography,etc.And it is widely present in our daily lives.With the improvement of spiritual and cultural pursuits,more and more people like to visit galleries,art exhibitions,etc.Therefore,it is important to help non-art professionals understand and appreciate visual arts.And many computer vision researchers focus on classifying and analyzing visual art images.Existing research usually uses deep learning methods to extract rich semantic features to classify visual art images.These methods mainly have the following two limitations.First,the styles of visual art images are closely related to art history,so only using visual features to classify images that ignores the impact of historical context on art development.Secondly,recognizing visual art images is different from general fine-grained classification tasks.Because the style of the artwork image does not only depend on the salient objects in the image,using conventional classification methods cannot distinguish the visual art style well.In view of the above problems,this paper starts with the multi-dimensional correlation to classify and analyze visual art images.Considering the connection between the development of painting styles and the background of art history,this paper explores the features of art history dimension and summarizes three factors that influence the formation and development of painting styles,i.e.,birthplace,origin time,and art movement.For different factors,designing knowledge extraction strategies respectively to generate label distributions.This method is encapsulated into an end-to-end multi-task learning framework.Experimental results show that the method is superior to single label methods that do not use art history dimension information.In addition,this paper proposes a convolutional neural network framework with adaptive cross-layer correlation,which captures texture information based on the visual dimension features to classify visual art images.This method weights features for different spatial locations and encapsulates them into an end-to-end network to learn the texture features of visual art images.This paper performs experiments on multiple visual art datasets to verify effectiveness and robustness.The experimental result of the proposed method outperforms existing style classification methods.