融合Trimap特征线索的人像抠图模型  

Portrait Image Matting Model Integrating Trimap Feature Clues

作　　者：姜宝石 谭小波[1] 张文波[1] 朱宏博 尹震宇 JIANG Baoshi;TAN Xiaobo;ZHANG Wenbo;ZHU Hongbo;YIN Zhenyu(School of Information Science and Engineering,Shenyang Ligong University,Shenyang 110159,China;School of Computer Science and Engineering,Northeastern University,Shenyang 110169,China;Shenyang Institute of Computing Technology,Chinese Academy of Sciences,Shenyang 110168,China)

机构地区：[1]沈阳理工大学信息科学与工程学院,沈阳110159 [2]东北大学计算机科学与工程学院,沈阳110169 [3]中国科学院沈阳计算技术研究所,沈阳110168

出　　处：《小型微型计算机系统》2025年第2期389-395,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金青年项目(62102272)资助。

摘　　要：近年来,基于深度学习的人像图像抠图技术已经成为研究热点.相比传统的抠图方法,深度学习抠图技术通过使用深度神经网络来学习和识别图像中的各种特征,从而进行精确的人像抠图,具有更高的精细度和稳定性.为了获得更精细的前景蒙版,Trimap常作为网络的额外的输入.然而,当前许多方法在利用Trimap的特征方面效率不高,并且都倾向于设计更深更复杂的模型网络,这可能会导致模型计算资源占用较大增加以及减缓模型的计算速度.针对以上情况,本文提出了一种以Trimap作为额外输入的简单架构抠图模型,具体来说,本文通过将Trimap图像与原始图像进行特征融合,使得抠图网络能够更好地关注到Trimap中的特征线索.此外,本文还设计了一种新型的归一化方式,以适应模型的训练过程.与近年流行的抠图模型相比,该模型在公开的人像数据集上显示出更低的量化损失(mse:0.0003,mad:0.0019,grad:6.0,conn:3.3),并且具有更精细的边界效果.In recent years,portrait image matting technology based on deep learning has emerged as a research hotspot.In comparison to traditional matting methods,deep learning matting techniques utilize deep neural networks to learn and recognize various features in images,enabling precise portrait matting with higher refinement and stability.To achieve a more refined foreground mask,Trimap is often used as an additional input to the network.However,many current methods and networks are inefficient in utilizing Trimap features,and they tend to design deeper and more complex model networks,potentially leading to increased model computational resource utilization and a slowdown in model computation speed.Addressing these challenges,this paper proposes a simple architecture matting model with Trimap as an additional input.Specifically,we fuse Trimap images with the original images to better focus the matting network on the features in the Trimap.Additionally,we introduce a novel normalization method designed to adapt to the model′s training process.Compared to popular matting models in recent years,our model demonstrates lower quantitative losses on public portrait datasets(mse:0.0003,mad:0.0019,grad:6.0,conn:3.3),and exhibits finer boundary effects.

关 键 词：数字图像抠图 深度学习技术 Trimap 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]