图像感知引导CycleGAN网络的背景虚化方法  

Background defocus method of image perception guided CycleGAN network

作　　者：叶武剑 林振溢 刘怡俊[1] 刘成民 YE Wu-jian;LIN Zhen-yi;LIU Yi-jun;LIU Cheng-min(School of Integrated Circuit,Guangdong University of Technology,Guangzhou 510006,China;School of Information Engineering,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学集成电路学院,广东广州510006 [2]广东工业大学信息工程学院,广东广州510006

出　　处：《液晶与显示》2023年第9期1248-1261,共14页Chinese Journal of Liquid Crystals and Displays

基　　金：广东省重点领域研发计划(No.2018B030338001,No.2018B010115002);2022年广州市基础研究计划基础与应用基础研究项目(No.202201010595);广东工业大学青年百人项目(No.220413548)。

摘　　要：现有生成对抗网络在背景虚化处理过程中,往往是无差别地提取整张输入图像的特征,导致网络难以区分图像的前后景,从而容易出现图像失真的现象。本文提出了图像感知引导CycleGAN网络的背景虚化方法,通过引入图像感知信息以提升模型性能。图像感知信息包括注意力信息和景深信息,前者用于引导网络关注不同的前后景区域,从而区分前后景;而后者用于增强前景目标的感知信息,能够实现有效的智能定焦并减少图像出现失真的现象,使背景虚化效果更佳。通过最小化生成对抗损失及循环一致性损失,可以避免丢失过多景深信息,提高图片的生成质量。实验结果及数据表明,提出的方法在背景虚化过程中能有效区分前后景并改善图像失真的现象,使生成的效果更加真实。此外,在与现有方法生成的图像效果对比中,通过采用问卷调查的方式进行了评估。本文提出的图像感知引导CycleGAN网络的背景虚化方法与SOTA相比,生成的图像质量更好,模型大小与生成图像的速率也具有明显的优势,分别为56.10 MB及47 ms。The existing image conversion algorithms based on generative adversarial network often extract the features of the whole input image indiscriminately in the process of background defocus,which makes it difficult for the network to distinguish the front and back scenes of the image,so it is easy to lead to the phenomenon of image distortion.We propose a background virtualization method of image perception guided CycleGAN network.The image perception information is introduced to improve the performance of the model.The image perception information includes attention information and depth of field information.The former is used to guide the network to pay attention to different foreground and background areas,so as to distinguish the foreground and background.The latter is used to enhance the perception information of foreground targets,achieve effective intelligent focusing,and reduce image distortion,making the background defocus better.The experimental results and data show that the method proposed in this paper can effectively distinguish the foreground and background in the process of background defocus,reduce the phenomenon of image distortion,and make the generated effect more real.In addition,in comparison with the image effect generated by the existing methods,the questionnaire survey is used for evaluation.A background virtualization method for image perception guided CycleGAN network is proposed,comparing with SOTA,the image quality generated is the best,and its model size and image generation rate also have obvious advantages of 56.10 MB and 47 ms,respectively.

关 键 词：背景虚化 图像感知 CycleGAN网络 智能定焦 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]