Extracting Generic Features of Artistic Style via Deep Convolutional Neural Network

While most existing works on image art style transformation generally focus on the transformation given a specific style image as input, in this paper, we consider it given a set of images of a generic style, e.g., images of Vincent van Gogh during 1889 to 1890. Compared to the specific style from only one input style image, our generic style transformation is able to remove the artifact generated from the single image such as specific objects and scenes. To this end, we propose a method to extract generic style features from a set of fine paintings. Generic style features describe these fine paintings from the global perspective, integrate features of brush strokes, color and pose contrast, scale information and orientation etc. We first obtain feature representation from these fine paintings using deep convolutional neural network (CNN), and then select generic representation from obtained representation. Finally, migrate visualized generic style features to input content image. Experimental results verify the efficiency and power of our method.