Evaluation of Painting Artistic Style Transfer Based on Generative Adversarial Network

Unlike most image style transfer tasks, which are to convert a photograph into an image with a certain artist’s style, this paper focuses on artistic style transfer in Monet painting, which is transferring a Monet painting into an image in another artist’s style. In our experiments, we successfully implemented the Cycle-Consistent GAN model and applied Neural Style Transfer (NST) model for contrasting effects. In order to evaluate the result of artistic style transfer quantitatively and effectively with a low requirement for computational resources, we also proposed a quantitative method called style transfer indicator to make the comparison more obvious as the comparisons of the effect of image style transfer were mostly done by subjective analysis previously. This method takes both the style and content of the transferred image into account because whether the transferred image belongs to the new style is as important as whether the content of the image is saved. A ResNet18 pre-trained model and structural similarity index are used for the evaluation of style and content respectively. The human survey that we conducted also proved the validity of our style transfer indicator. Moreover, our proposed indicator could also be applied for the evaluation of other image style transfer tasks.