Stroke Controllable Fast Style Transfer with Adaptive Receptive Fields.

Recently, in the community of Neural Style Transfer, several algorithms areproposed to transfer an artistic style in real-time, which is known as FastStyle Transfer. However, controlling the stroke size in stylized results stillremains an open challenge. To achieve controllable stroke sizes, severalattempts were made including training multiple models and resizing the inputimage in a variety of scales, respectively. However, their results are notpromising regarding the efficiency and quality. In this paper, we present astroke controllable style transfer network that incorporates different strokesizes into one single model. Firstly, by analyzing the factors that influencethe stroke size, we adopt the idea that both the receptive field and the styleimage scale should be taken into consideration for most cases. Then we proposea StrokePyramid module to endow the network with adaptive receptive fields, andtwo training strategies to achieve faster convergence and augment new strokesizes upon a trained model respectively. Finally, by combining the proposedruntime control techniques, our network can produce distinct stroke sizes indifferent output images or different spatial regions within the same outputimage. The experimental results demonstrate that with almost the same number ofparameters as the previous Fast Style Transfer algorithm, our network cantransfer an artistic style in a stroke controllable manner.