[1]解奕鹏,秦品乐,曾建潮,等.基于姿态注意力的特定角度人脸正面化网络[J].计算机技术与发展,2023,33(07):47-54.[doi:10. 3969 / j. issn. 1673-629X. 2023. 07. 007]
　XIE Yi-peng,QIN Pin-le,ZENG Jian-chao,et al.Face Frontalization Network of Specific Angle Based on Pose Attention[J].,2023,33(07):47-54.[doi:10. 3969 / j. issn. 1673-629X. 2023. 07. 007]
点击复制

基于姿态注意力的特定角度人脸正面化网络()
分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:
33
期数:
2023年07期
页码:
47-54
栏目:
媒体计算
出版日期:
2023-07-10

文章信息/Info

Title:
Face Frontalization Network of Specific Angle Based on Pose Attention
文章编号:
1673-629X(2023)07-0047-08
作者:
解奕鹏1; 2 ; 秦品乐1; 2 ; 曾建潮1; 2 ; 闫寒梅3 ; 柴 锐1; 2 ; 赵鹏程1; 2
1. 中北大学 大数据学院,山西 太原 030051;
2. 山西省医学影像与数据分析工程研究中心(中北大学),山西 太原 030051;
3. 山西警察学院 刑事科学技术系,山西 太原 030401
Author(s):
XIE Yi-peng1; 2 ; QIN Pin-le1; 2 ; ZENG Jian-chao1; 2 ; YAN Han-mei3 ; CHAI Rui1; 2 ; ZHAO Peng-cheng1; 2
1. School of Big Data,North University of China,Taiyuan 030051,China;
2. Shanxi Medical Imaging and Data Analysis Engineering Research Center ( North University of China) ,Taiyuan 030051,China;
3. Criminal Science and Technology,Shanxi Police College,Taiyuan 030401,China
关键词:
人脸正面化; 注意力机制; 生成对抗网络; 人脸识别; 深度学习
Keywords:
face frontalization; attention mechanism; generative adversarial network ( GAN) ; face recognition; deep learning
分类号:
TP391. 41
DOI:
10. 3969 / j. issn. 1673-629X. 2023. 07. 007
摘要:
人脸正面化对人脸识别有重要意义,但实际监控场景中大姿态的人脸正面化效果通常不如较小姿态,因此提出姿态引导的特定角度生成对抗网络( Pose-Specific Generative Adversarial Network,PS-GAN) 。 PS-GAN 由生成器和鉴别器组成,生成器由编码器、姿态注意模块、特征转换模块以及解码器四部分组成,编码器与解码器分别对输入图像进行下采样与上采样,姿态注意模块为网络引入人脸结构先验的同时约束模型关注感兴趣区域,特征转换模块对编码器得到的侧脸特征进行变换并抑制冗余通道。 首先,将连续的姿态变化划分为离散的姿态集合,单个 PS-GAN 模型由某一特定角度的数据训练;然后,将多个 PS-GAN 进行组合,使其适用于任意角度的人脸输入。 在本实验室自主采集的 MASFD 数据集以及 CAS-PEAL-R1 公开数据集上进行了大量的定性与定量实验,验证了网络结构的有效性以及合理性;与现有方法相比,虽然 PS-GAN 是由受限数据集训练的,但它也能在非同源数据上有良好的视觉效果。
Abstract:
Face frontalization is of great significance for face recognition,but in actual monitoring scenarios,the reconstruction results oflarge postures is usually not as good as that of small postures. Therefore,we propose a Pose-Specific Generative Adversarial Network(PS-GAN) ,which consists of a generator and a discriminator. The generator consists of an encoder,a pose attention module,a featureconversion module,and a decoder. The encoder and decoder downsample and upsample the input image,respectively. The pose attentionmodule introduces face structure prior into the network and simultaneously constrains the model to focus on the region of interest. Thefeature transformation module transforms the profile features obtained by the encoder and suppresses redundant channels. Firstly,the continuous pose are divided into discrete pose sets. A single PS-GAN model is trained by data from a specific Angle,and then multiple PS-GANs are combined to make it suitable for face input from any Angle. In addition, a large number of qualitative and quantitative　experiments were carried out on the MASFD data set independently collected by our laboratory and the CAS-PEAL-R1 public data set,which verified the validity and rationality of the network structure. Compared with existing methods,　although PS-GAN is trained on arestricted dataset,but it also has excellent visual performance on non-homologous data.

相似文献/References:

[1]李梦洁,董峦.基于 PyTorch 的机器翻译算法的实现[J].计算机技术与发展,2018,28(10):160.[doi:10.3969/ j. issn.1673-629X.2018.10.033]
　LI Meng-jie,DONG Luan.Implementation of Machine Translation Algorithm Based on PyTorch[J].,2018,28(07):160.[doi:10.3969/ j. issn.1673-629X.2018.10.033]
[2]李东欣,禹 龙,田生伟,等.注意力机制的 LSTM-DBN 维语人称代词指代消解[J].计算机技术与发展,2019,29(07):33.[doi:10. 3969 / j. issn. 1673-629X. 2019. 07. 007]
　LI Dong-xin,YU Long,TIAN Sheng-wei,et al.Attention Mechanism of LSTM-DBN Uyghur Personal Pronoun Anaphora Resolution[J].,2019,29(07):33.[doi:10. 3969 / j. issn. 1673-629X. 2019. 07. 007]
[3]尹 鹏,周 林,郭 强,等.基于短语级注意力机制的关系抽取方法[J].计算机技术与发展,2019,29(09):24.[doi:10. 3969 / j. issn. 1673-629X. 2019. 09. 005]
　YIN Peng,ZHOU Lin,GUO Qiang,et al.Relation Extraction Based on Phrase-level Attention[J].,2019,29(07):24.[doi:10. 3969 / j. issn. 1673-629X. 2019. 09. 005]
[4]钟 诚,周浩杰,韦海亮.一种基于注意力机制的三维点云物体识别方法[J].计算机技术与发展,2020,30(04):41.[doi:10. 3969 / j. issn. 1673-629X. 2020. 04. 008]
　ZHONG Cheng,ZHOU Hao-jie,WEI Hai-liang.A 3D Point Cloud Object Recognition Method Based on Attention Mechanism[J].,2020,30(07):41.[doi:10. 3969 / j. issn. 1673-629X. 2020. 04. 008]
[5]王乾铭,李 吟.基于深度学习的个性化聊天机器人研究[J].计算机技术与发展,2020,30(04):79.[doi:10. 3969 / j. issn. 1673-629X. 2020. 04. 015]
　WANG Qian-ming,LI Yin.Research on Personalized Chatbot Based on Deep Learning[J].,2020,30(07):79.[doi:10. 3969 / j. issn. 1673-629X. 2020. 04. 015]
[6]朱立倩.基于深度学习的数显仪表字符识别[J].计算机技术与发展,2020,30(06):141.[doi:10. 3969 / j. issn. 1673-629X. 2020. 06. 027]
　ZHU Li-qian.Character Recognition of Digital Display Instrument Based on Deep Learning[J].,2020,30(07):141.[doi:10. 3969 / j. issn. 1673-629X. 2020. 06. 027]
[7]申静波,李井辉,孙丽娜.注意力机制在评论文本情感分析中的应用研究[J].计算机技术与发展,2020,30(07):169.[doi:10. 3969 / j. issn. 1673-629X. 2020. 07. 036]
　SHEN Jing-bo,LI Jing-hui,SUN Li-na.Research on Application of Attention Mechanism in Comment Text Emotional Analysis[J].,2020,30(07):169.[doi:10. 3969 / j. issn. 1673-629X. 2020. 07. 036]
[8]张亚飞.基于注意力的权重分配机制[J].计算机技术与发展,2020,30(09):49.[doi:10. 3969 / j. issn. 1673-629X. 2020. 09. 009]
　ZHANG Ya-fei.Attention-based Weight Allocation Mechanism[J].,2020,30(07):49.[doi:10. 3969 / j. issn. 1673-629X. 2020. 09. 009]
[9]王振业,叶成绪*,王文韬,等.基于 LSTM-Att 方法的音乐流行趋势预测[J].计算机技术与发展,2020,30(09):188.[doi:10. 3969 / j. issn. 1673-629X. 2020. 09. 034]
　WANG Zhen-ye,YE Cheng-xu*,WANG Wen-tao,et al.Music Trend Forecast Based on LSTM-Att Method[J].,2020,30(07):188.[doi:10. 3969 / j. issn. 1673-629X. 2020. 09. 034]
[10]陈 琛,刘小云,方玉华.融合注意力机制的电子病历命名实体识别[J].计算机技术与发展,2020,30(10):216.[doi:10. 3969 / j. issn. 1673-629X. 2020. 10. 038]
　CHEN Chen,LIU Xiao-yun,FANG Yu-hua.Named Entity Recognition in Electronic Medical Record Introducing Attention Mechanisms[J].,2020,30(07):216.[doi:10. 3969 / j. issn. 1673-629X. 2020. 10. 038]

更新日期/Last Update: 2023-07-10