摘要
将图像风格迁移技术引入家居风格设计领域。传统图像风格迁移方法操作复杂且迁移效果很差,深度学习方法虽在效果上有了很大提高,但是迁移效果偏于艺术化而导致图像失真,而且在迁移过程中容易发生图像内容迁移错误。针对这些情况,基于图像梯度约束对家居设计的风格迁移进行研究。用图像分割技术将风格迁移局限在相同语义内容的区域;通过Gram矩阵计算纹理特征统计相关性为每个语义类别构造单独的风格损失;通过泊松图像编辑方法对风格化图像进行梯度约束。实验结果表明,该方法避免了出现迁移内容错误以及畸变且失真的问题,取得较好的迁移效果图,可以实际应用于室内装饰中。
This paper proposes to introduce the image style transfer technology into the field of home style design.The traditional image style transfer method is complicated and the transfer effect is very poor.The deep learning method has greatly improved in effect,but the transfer effect is partial to artistic,which leads to image distortion.Moreover,image content transfer errors and distortions are likely to occur during the transfer process.In view of these situations,the style transfer of home design is studied based on the image gradient constraint.The image segmentation technology was used to limit the style transfer to the area of the same semantic content;the statistical correlation of texture features was calculated by Gram matrix to construct a separate style loss for each semantic category;the Poisson image editing method was used to constrain the gradient of stylized image.The experimental results show that the method avoids the problems of transfer content error and distortion.And it obtains a better transfer effect map,which can be practically applied in interior decoration.
作者
冯威
诸跃进
肖金球
段杰
周惟
Feng Wei;Zhu Yuejin;Xiao Jinqiu;Duan Jie;Zhou Wei(College of Electronics and Information Engineering,Suzhou University of Science and Technology,Suzhou 215009,Jiangsu,China)
出处
《计算机应用与软件》
北大核心
2020年第7期170-175,245,共7页
Computer Applications and Software
关键词
室内装饰
深度学习
风格迁移
家居风格
泊松图像编辑
Interior decoration
Deep learning
Style transfer
Home style
Poisson image editing