摘要
针对复杂背景下服装图像局部区域风格迁移难以控制及迁移后容易产生边界伪影的问题,提出一种基于注意力机制的CycleGAN服装局部风格迁移方法。通过VGG16网络分别提取服装图像的内容特征与风格特征,将其输入基于注意力机制的CycleGAN生成器中,应用注意力机制在复杂背景下的各个服装区域分配概率分布信息,获得注意力分布更多的区域及相关度更高的区域,并采用改进的损失函数校正边界伪影,对该区域进行风格迁移得到所需的风格迁移服装图像。实验结果表明,与CNN、FCN、BeautyGAN图像局部风格迁移方法相比,该方法不仅可以突出服装图像局部风格迁移效果,而且增强了图像细节,有利于提高输出图像的真实性和艺术性。
For the clothing images with complex backgrounds,it is difficult to control the style transfer in local areas of the images,and boundary artifacts are easily generated in this process.To address the problem,a CycleGAN-based method using attention mechanism is proposed for local style transfer of clothing images.The method employs VGG16 to extract the content features and style features from the clothing images separately,and the features are subsequently input into the CycleGAN generator based on the attention mechanism.The attention mechanism is used to distribute the probability distribution information to each clothing area of the complex background to obtain the area with more attention distribution and the area with higher correlation.Then an improved loss function is used to correct boundary artifacts.Finally,style transfer is performed on the area to obtain the desired style transfer clothing image.The experimental results show that compared with the local style transfer methods based on CNN,FCN and BeautyGAN,the proposed method can not only highlight the effect of local style transfer of clothing images,but also enhance the image details,which is conducive to improving the authenticity and artistry of the output images.
作者
陈佳
董学良
梁金星
何儒汉
CHEN Jia;DONG Xueliang;LIANG Jinxing;HE Ruhan(School of Mathematics and Computer Science,Wuhan Textile University,Wuhan 430000,China;Engineering Research Center of Hubei Province for Clothing Information,Wuhan 430000,China)
出处
《计算机工程》
CAS
CSCD
北大核心
2021年第11期305-312,共8页
Computer Engineering
基金
湖北省自然科学基金计划一般面上项目(2020CFB801)
湖北省高等学校优秀中青年科技创新团队计划(T201807)。
关键词
图像风格迁移
边界伪影
注意力机制
循环生成对抗网络
损失函数
image style transfer
boundary artifacts
attention mechanism
Cycle Generative Adversarial Network(CycleGAN)
loss function
