Creation of art is a complex process for its abstraction and novelty. In order to create those art with less cost, style transfer using advanced machine learning technology becomes a popular method in computer vision field. However, traditional transferred image still troubles with color anamorphosis, content losing, and time-consuming problems. In this paper, we propose an improved style transfer algorithm using the feedforward neural network. The whole network is composed of two parts, a style transfer network and a loss network. The style transfer network owns the ability of directly mapping the content image into the stylized image after training. Content loss, style loss, and Total Variation (TV) loss are calculated by the loss network to update the weight of the style transfer network. Additionally, a cross training strategy is proposed to better preserve the details of the content image. Plenty of experiments are conducted to show the superior performance of our presented algorithm compared to the classic neural style transfer algorithm.
Human-robot cross-training: Computational formulation, modeling and evaluation of a human team training strategy