摘要
视频制作过程中,前景与背景的不一致性严重降低了合成图像的质量,视频制作系统通过调节一系列参数选项才能使前景与背景相协调,大大提高了视频制作的成本。针对合成图像和谐化后前景与背景不相融,色彩失真等问题,基于生成对抗网络,采用跨尺度跳层连接具有轻量化网络结构,建立图像合成模型。应用生成对抗网络的“域验证”思想,在生成器部分引入跨尺度跳层连接,设计色差一致性损失函数,引入多用卷积核等优化图像和谐化流程。结果表明,上述模型有更轻量化的网络结构和更好的人像和谐化效果。
In the process of video production,the inconsistency between foreground and background seriously reduces the quality of the synthesized image.The video production system can coordinate foreground and background by adjusting a series of parameter options,which greatly increases the cost of video production.In order to solve the problems,such as incompatibility between foreground and background and color distortion after harmonization of synthesized images,this paper establishes an image synthesis model based on a generation adversarial network and cross-scale skipping layer connection with a lightweight network structure.By using the idea of"domain verification"of the generated adversarial network,a cross-scale skip layer connection is introduced into the generator,a color difference consistency loss function is designed,and a multi-purpose convolution kernel is introduced to optimize the image harmonization process.The results show that this model has a lighter network structure and a better portrait harmonization effect.
作者
翟莹莹
尹泓霁
厉英
孙巍
ZHAI Ying-ying;YIN Hong-ji;LI Ying;SUN Wei(Liaoning Zhongwang Group Co.,Ltd,Liaoyang Liaoning 111003,China;College of Computer Science and Engineering,Northeastern University,Shenyang Liaoning 110169,China)
出处
《计算机仿真》
2025年第10期536-541,共6页
Computer Simulation
基金
中央高校基本科研业务费资助项目(N2216011)
国家自然科学基金项目(62232007)。
关键词
图像合成
人像和谐化
生成对抗网络
域验证
Image synthesis
Portrait harmonization
Generative adversarial networks
Domain verification
