SIA OpenIR  > 其他
Alternative TitleResearch on Face Recognition of Intelligent Service Robot Based on Deep Learning
Thesis Advisor王宏玉
Keyword深度学习 人脸识别 人脸检测 人脸对齐 智能服务机器人
Call NumberTP391.41/Z36/2018
Degree Discipline控制工程
Degree Name硕士
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract本文以国家科技支撑计划“国产机器人嵌入式实时操作系统开发与应用示范”为依托,针对目前人脸识别算法在复杂实际应用场景中识别率较低,鲁棒性较差,难以本地实时运算等问题,以及目前新松公司智能服务机器人人脸识别性能指标需要进一步提高,应用领域需要进一步拓展的需求。本课题面向新松智能服务机器人实际应用场景,基于人工智能中的深度学习方法,围绕人脸检测、人脸对齐和人脸识别三个关键模块,进行深入的理论分析与实际应用研究,提出一种具有较强鲁棒性的高精度实时人脸识别方法,并基于NVIDIA TX2嵌入式板卡实现基于深度学习的智能服务机器人人脸识别系统。本文的主要研究成果有:(1)在人脸检测和人脸对齐部分,提出了一种新的级联多任务卷积神经网络,同时实现了人脸检测和面部特征点定位任务。该方法将传统方法中的级联模式引入到深度卷积神经网络,由粗到精对候选区域进行判别。同时,多任务学习将人脸检测和面部特征点定位两个任务融合到一个深度框架中,不仅提高了算法的效率,而且可以充分挖掘任务之间的相关性,提高检测准确率。此外,在第一级模型末端采用了新的分块损失函数,使模型训练时更好地关注局部信息,对由于遮挡和姿态变化产生的部分信息损失具有较好的鲁棒性。最终,该方法在FDDB人脸检测评测数据集上100个误检时达到了90%准确率,在AFLW评测数据集上面部特征点平均定位误差优于传统算法。(2)在人脸识别部分,首先,利用通用三维人脸模板进行多姿态人脸生成,从训练数据姿态增广角度为人脸识别中的大姿态问题提供了一种有效的解决方式。然后,针对深度人脸识别模型参数量巨大的问题,提出了一种轻量化的NIN-ResNet模型。同时,设计了中心损失函数来对类内距离进行度量学习,并以分类损失与中心损失作为模型训练的联合监督信号,增强模型的判别能力,提高人脸识别的鲁棒性和准确性。最终,识别准确率在人脸识别评测数据集LFW和YouTube Faces分别达到了99.03%和92.1%。(3)在智能服务机器人人脸识别系统部署部分,由于NVIDIA TX2嵌入式板卡带有高性能的GPU模块,在深度学习算法加速与优化方面具有很大的优势。因此本文提出了基于NVIDIA TX2嵌入式板卡并采用Caffe开源深度学习框架来进行智能服务机器人人脸识别系统部署的方案。同时,为了更好的验证系统部署到智能服务机器人上的整体性能,我们在新松智能服务机器人上进行人脸数据采集,构建了实际应用场景中的人脸识别评测数据集。在该数据集上我们的人脸识别算法达到了99.22%的准确率,人脸检测、人脸对齐、人脸识别整套系统流程在80ms左右。
Other AbstractThis article is based on the national science and technology support program "the development and application demonstration of the embedded real-time operating system for domestic robot". Aiming at the face recognition algorithm in complex application scenarios in the low recognition rate, poor robustness, poor real-time, and the performance of SIASUN's intelligent service robot face recognition needs to be further improved, applications areas need to be further expanded. The subject is oriented to the practical application scene of SIASUN intelligent service robot, based on the deep learning method in artificial intelligence. Based on three key modules, face detection, face alignment and face recognition, in-depth theoretical analysis and practical application are carried out. A high-accuracy face recognition method with strong robustness and real time performance is proposed, and the intelligent service robot face recognition system based on deep learning is implemented based on NVIDIA TX2 embedded card. The main research results of this paper are as follows: (1) In face detection and face alignment part, a new cascaded multitask convolutional neural network is proposed. Face detection and facial feature points location tasks are implemented simultaneously. This method introduces the cascade mode in the traditional method to the deep convolutional neural network, and discriminating the candidate regions from coarse to fine. At the same time, multitask learning integrates two tasks of face detection and facial feature location into one deep framework, which not only improves the efficiency of the algorithm, but also fully explores the correlation between tasks and improves the accuracy of detection. In addition, a new block loss function is adopted at the end of the first level model, which makes the model training better focus on local information and has better robustness to some information loss due to occlusion and posture changes. Finally, the algorithm achieves 90% accuracy in 100 erroneous detection on FDDB face detection and evaluation dataset, and the average location error of facial feature points is better than the traditional algorithm on AFLW evaluation dataset. (2) In the face recognition part, first of all, we use universal 3D face template to generate multi pose face. It provides an effective solution for the big pose problem in face recognition from the perspective of training data augmentation. Then, a lightweight NIN-ResNet model is proposed to solve the problem of the huge parameter of the deep face recognition model. At the same time, a central loss function is designed to measure the intra class distance, and the classification loss and center loss are used as joint monitoring signals for model training, which enhance the discriminant ability of the model and improve the robustness and accuracy of face recognition. In the end, the recognition accuracy was 99.03% and 92.1% in the face recognition evaluation data set LFW and YouTube Faces respectively. (3) In the deployment part of the intelligent service robot face recognition system, because NVIDIA TX2 embedded card has high performance GPU module, it has great advantages in accelerating and optimizing deep learning algorithm. Therefore, this paper proposes a scheme for the deployment of intelligent service robot face recognition system based on NVIDIA TX2 embedded card and Caffe open source deep learning framework. At the same time, in order to better verify the overall performance of the system to the intelligent service robot, we collect the face data on the SIASUN intelligent service robot, and build the face recognition evaluation data set in the actual application scene. On this data set, our face recognition algorithm has reached 99.22% accuracy, and the whole system process of face detection, face alignment and face recognition is around 80ms.
Contribution Rank1
Document Type学位论文
Recommended Citation
GB/T 7714
张延安. 基于深度学习的智能服务机器人人脸识别研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2018.
Files in This Item:
File Name/Size DocType Version Access License
基于深度学习的智能服务机器人人脸识别研究(4346KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[张延安]'s Articles
Baidu academic
Similar articles in Baidu academic
[张延安]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[张延安]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.