Research On Painting Image Classification Based On Aesthetic Style

Painting is an important expression of culture and art in the development of human civilization, and over the years a large number of paintings have been produced. Therefore, researching on such paintings is treated as a critical tool for people to understand human history, culture, art, and technology development history, which could further promote the prosperity of human civilization. With the development and wide application of digital technology, more and more paintings gradually digitized, making large-scale paintings art analysis possible. However, researchers are facing new problems of how to effectively use these massive digital resources when a large number of digitized paintings bring a wealth of image resources for them. Consequently, how to ultilize the computer to classify these massive painting images to facilitate further study of the researchers is recognized as an important issue.For artificial works, there exists a deep gap between paintings and natural images, due to special characteristics of paintings. Therefore, general image classification methods are not suitable to be applied directly to the paintings. In this paper, we attempt to utilize special aesthetic style characteristics of paintings to accomplish classification tasks among massive digited painting images.The main contributions of this paper include:Firstly, based on painting techniques, we propose a method to construct the descriptor of aesthetic style. After careful studies on art literature and research on the Chinese-Western paintings and different dynasties Dunhuang murals, we sum three features or three categories of properties, sixteen features to describe the aesthetic style for Chinese-Western paintings or Dunhuang murals that lie in different dynasties respectively from the view of painting technology, and the classification of paintings has been achieved. Classification experiments have validated the proposed idea.Secondly, we present an aesthetic style classification method based on the aesthetic style similarity. By researching on cognitive mechanisms of the human brain and the principle of similarity, digging the art style attributes, we establish the aesthetic style similarity rules. Then, built upon on above rules, we quantify the aesthetic features that generally accepted in the art domain, calculate self-similarity descriptors of image-style, then compute similarity coefficients between the image and each other images to constitute the similarity matrix, and finally judge the unknown samples by using of the Adaboost algorithm. The experimental results also proved the efficiency of using the aesthetic style similarity rules for painting images classification.Finally, we propose a method for aesthetic style classification based on saliency map. Studies have shown that the human visual sensory has a selective visual attention on the information presented by the world. Firstly, the color enhancement algorithms and the global region contrast based algorithm are used to calculate the saliency map of images. Secondly, we classify images based on probability model by using of the computed saliency map. Experimental results on the different aesthetic style Dunhuang murals database, as well as two other common classification databases like Caltech101and Caltech256show that, the proposed method not only can be successfully applied to tasks of the aesthetic style classification of painting images, but also can be extended to generic image classification tasks, indicating the strong robustness of the method.