Image neural style transfer combining global and local optimization

In order to avoid the shortcomings of a single optimization method, improve the effect of style transfer, and control the occurrence of artifacts, this paper proposes a neural style transfer method combining global and local optimization. In the calculation of local loss, the content mask and style mask are used to the matching process of the image patches to preserve the style details and reduce the mismatching of the image. The global loss function is calculated by Gram matrix, and the mask of the content feature map is added to the feature map of the synthetic image. The effect of mask data on the image is controlled by hyperparameters, and the Laplacian operator is introduced for structural refinement to better preserve the structural integrity of the stylized image. Experimental results show that this method can extend the scope of application of style transfer, can be effectively used for different images, and effectively control artifacts. The data for our approach are publicly available at https://github.com/xlyusegithub/styletransfer.git.