高级检索
李应涛, 徐丹. 木刻版画风格转换的深度学习算法[J]. 计算机辅助设计与图形学学报, 2020, 32(11): 1804-1812. doi: 10.3724/SP.J.1089.2020.18148
引用本文: 李应涛, 徐丹. 木刻版画风格转换的深度学习算法[J]. 计算机辅助设计与图形学学报, 2020, 32(11): 1804-1812. doi: 10.3724/SP.J.1089.2020.18148
Li Yingtao, Xu Dan. Deep Learning Algorithm for Woodcut Prints Style Transfer[J]. Journal of Computer-Aided Design & Computer Graphics, 2020, 32(11): 1804-1812. doi: 10.3724/SP.J.1089.2020.18148
Citation: Li Yingtao, Xu Dan. Deep Learning Algorithm for Woodcut Prints Style Transfer[J]. Journal of Computer-Aided Design & Computer Graphics, 2020, 32(11): 1804-1812. doi: 10.3724/SP.J.1089.2020.18148

木刻版画风格转换的深度学习算法

Deep Learning Algorithm for Woodcut Prints Style Transfer

  • 摘要: 为了使木刻版画风格转换结果呈现出更明显的木刻刻痕纹理,同时保持刻痕纹理分布的合理性,提出一种基于神经网络语义分割算法和神经风格转换的木刻版画风格转换算法,该算法按不同区域进行木刻版画的风格转换.首先,使用神经网络分割算法和Labelme图像标注工具分别对内容图像和木刻版画图像进行语义分割.然后将分割结果二值化,形成掩膜图像.将掩膜图像作为引导,与内容图像和木刻版画图像一起输入具有空间引导通道的神经风格转换网络进行分区域风格转换.在PyTorch深度学习框架下,使用该算法对大量人物和自然场景图片进行木刻版画风格转换,并与基于迭代优化、快速风格转换和任意风格转换3类神经风格转换算法中各自最具代表性算法的转换结果进行比较.结果表明,所提算法的木刻版画风格转换结果所呈现的木刻刻痕纹理明显,刻痕纹理分布合理,转换结果真实自然,更接近真实的木刻版画.

     

    Abstract: In order to make the style transferred result of woodcut prints shows a more obvious woodcut nick texture,while maintaining the rationality of the nick texture distribution,a woodcut prints style transfer algorithm based on neural network semantic segmentation algorithm and neural style transfer is proposed,which transfers woodcut prints style according to different regions.Firstly,neural network segmentation algorithm and Labelme image annotation tool are used to segment the content image and the woodcut prints image respectively.Then binarization process the segmentation results to formed mask images.The mask images are used as guidances,and input into the neural style transfer network with spatial guidance channels for regional style transfer together with the content image and the woodcut prints image.Under the deep learning framework of PyTorch,the algorithm is used to transfer a large number of images of characters and natural scenes into woodcut prints,and the transferred results are compared with those of the state-of-art neural style transfer algorithms including iterative optimization,fast style transfer and arbitrary style transfer.Experiments show that the proposed woodcut prints style transfer algorithm has more obvious woodcut nick texture,and the nick texture distribution is more reasonable;and also the transferred results are more real and natural,which are closer to real woodcut prints.

     

/

返回文章
返回