Abstract: Image style transfer is a research hot topic in computer graphics and computer vision.Aiming at the difficulty in the style transfer of the local area of the content image in the existing image style transfer method,this paper proposed a localized image transfer framework based on convolutional neural network.First,according to the input content image and style image,the image style transfer network is used to generate the whole style transferred image.Then,the image foreground and the background area are determined by the mask generated by automatic semantic segmentation.Finally,according to style transfer result of the foreground or the background region,an image fusion algorithm based on Manhattan distance is proposed to optimize the convergence and smooth transition between the stylized object and the original area.The framework comprehensively considers the pixel values and positions of the target area and the boundary band,and experiments on three public image datasets demonstrate that the method can efficiently,quickly and naturally implement local style transfer of input content maps,and produce visual effects that are both artistic and authentic.

Key words: Automatic semantic segmentation, Convolutional neural network (CNN), Deep learning, Localized image style transfer, Manhattan distance