关键词: 风格迁移, 卷积神经网络, 语义分割, 区域差异

Abstract:

Image style transfer techniques always can be used to automatically endow different styles to an image based on a given style image. Only the whole image or a single area in the image used to be stylized through the existing image style transfer method, which inevitable results in limitations in practical applications. By introducing the semantic information of the content image in the style conversion process, the different semantic regions of the image can be stylized differently. The content image is semantically segmented and introduced into the VGG loss network to define the stylized region of the image. The Gram matrices are respectively calculated on each of the regions, and gradient propagation is limited to each semantic region in the back-propagation stage to obtain the style feature values for different regions of the image. The regularization loss is introduced into the loss function, which aims to reduce the interaction between different regions. Experiments are designed on the Microsoft COCO2017 dataset, and the results show that the method has achieved different stylization of multiple regions of the image while ensuring the natural transition between regions.

Key words: style transfer, convolutional neural network, semantic segmentation, regional differences