Style-oriented representative paintings selection

Style transfer is used as a means to render an image in the artistic style of another one. An ideal style transfer algorithm should be able to extract and represent the semantic image content from the source image and then render the content in the style of the example image. The decomposition of content and style in artistic images is bound to the coupling between the source content and the example style. Previous image style transfer works only focus on expressing the artistic style of a specific painting [Gatys et al. 2016; Liao et al. 2017]. However, the painting styles of an artist may vary throughout all his/her painting works. Usually we need to find multiple art works to represent an artist's creation characteristics, so that we can generate a series of stylized images with the artist's painting styles. In this work, we proposed a novel method to select representative paintings of an artist. Different from traditional clustering problems, we don't try to assign each image a correct label. We focus on finding the most representative ones in all the paintings. We first use K -means to preliminary cluster an artist's paintings. Clustering centres are the original representative images. Then we employ rejection to pick out the unrepresentative and confused samples. Finally, we update the K classes and get the new representative images.