The digitization of traditional Chinese painting (TCP) is important for its preserving and sharing. Massive traditional Chinese painting has become an important part of the digital museum. How to achieve efficient retrieval and management of traditional Chinese paintings has become an urgent demands. There are two categories of approaches to handle this problem. The first category is methods based on manual annotation, trying to retrieval TCPs using the tags that labeled by annotators for each image. This kind of methods can get promising results for small dataset. But when the scale of dataset is large, these methods need a large number of annotators for manual annotation and a lot of time. This cannot meet the needs of the rapid growth in the number of TCPs. Furthermore, manual annotations are subjective, different annotators may have different understanding for the same traditional Chinese painting. The second category is methods based on image content which analyses the content of the traditional Chinese painting automatically. With the development of the theory of image analysis and retrieval, this kind of methods are widely concerned and researched. On contrary to the first kind of methods, there are many merits, e.g., they are objective and they can deal with large-scale dataset. The proposed method belongs to the latter category and focus on the classification of TCP. In this paper, we proposed a novel traditional Chinese painting classification method based on painting techniques. Generally, according to the differences of painting techniques, the traditional Chinese painting can be divided into two categories, i.e., Gongbi and Xieyi. Due to various contents, different art styles of different painters, classification of Gongbi and Xieyi is still a challenging task. By analyzing the property of Gongbi and Xieyi, we proposed a novel key region detection method for TCP firstly. The main differences between 'Gongbi' and 'Xieyi' are the representation of details. In general, Gongbi paints the outline first, then coloring, Xieyi is not. Therefore, key regions for districting Gongbi and Xieyi should consist of point and edge. According to this, in this step, we incorporated the Scale Invariant Feature Transform (SIFT) detector and canny edge detector for detecting point and edge respectively. The points and edges obtained constitute the key regions for distinguishing Gongbi and Xieyi. Secondly, we proposed a Neighborhood Difference Descriptor (NDD). NDD represents the similarity and difference for a neighborhood of each edge pixel. The final descriptors consisted of the NDD and Bag of words. Finally, a cascade classification scheme was introduced, which combined the different dimension features to obtain the classification results. Comprehensive and comparative experiments were done over a challenging dataset. The dataset consists of 1718 traditional Chinese paintings and the numbers of Gongbi and Xieyi are 822 and 896 respectively. We split the dataset into training set and testing set. The key parameter and the classification were trained on the training set. The experimental results showed that the proposed feature can describe the painting techniques property effectively and achieved promising performance. Comparative experiments showed that our method outperforms the state-of-the-art methods. © 2017, Science Press. All right reserved.
