Stroke Controllable Fast Style Transfer with Adaptive Receptive Fields
share
Recently, in the community of Neural Style Transfer, several algorithms are proposed to transfer an artistic style in real-time, which is known as Fast Style Transfer. However, controlling the stroke size in stylized results still remains an open challenge. To achieve controllable stroke sizes, several attempts were made including training multiple models and resizing the input image in a variety of scales, respectively. However, their results are not promising regarding the efficiency and quality. In this paper, we present a stroke controllable style transfer network that incorporates different stroke sizes into one single model. Firstly, by analyzing the factors that influence the stroke size, we adopt the idea that both the receptive field and the style image scale should be taken into consideration for most cases. Then we propose a StrokePyramid module to endow the network with adaptive receptive fields, and two training strategies to achieve faster convergence and augment new stroke sizes upon a trained model respectively. Finally, by combining the proposed runtime control techniques, our network can produce distinct stroke sizes in different output images or different spatial regions within the same output image. The experimental results demonstrate that with almost the same number of parameters as the previous Fast Style Transfer algorithm, our network can transfer an artistic style in a stroke controllable manner.
READ FULL TEXT
