CASIA OpenIR  > 毕业生  > 博士学位论文
基于深度学习的人脸及行人识别算法研究
石海林
学位类型工学博士
导师李子青
2016-05-26
学位授予单位中国科学院研究生院
学位授予地点北京
关键词人脸识别 行人再辨识 深度学习 监控图像 异质图像匹配
摘要
人脸与行人识别是生物识别与安全技术领域中的关键技术。
实际应用中,由于光照、姿态、模糊、低分辨率、遮挡等因素的影响,人脸与行人数据往往包含着多种噪声,并由此带来较大的类内差异,给识别性能造成重大影响。
本论文的主要工作集中于解决人脸与行人识别中存在的这些困难问题。具体来说,本文主要对基于深度学习的人脸识别算法和行人再辨识算法进行改进并提出相关新方法,其主要贡献点可归纳为以下几点:

1 基于自编码算法,提出了将自我重建的训练目标推广至同类样本之间的相互重建,称之为类编码算法。该算法为网络的训练提供了高维特征空间中更有效的类内约束。此外,还将类内的相互重建从数据层改进至特征层,从而使得训练过程对图像噪声和平移具有鲁棒性;并与深度卷积神经网络结合组成深度类编码网络,从而学习到兼具鲁棒性和鉴别性的人脸特征,用于非限制人脸识别。
 
2  针对监控视频条件下人脸识别中存在的人脸图像异质性以及标记的训练数据量太小的问题,提出了基于多网络多任务的深度学习算法用于监控人脸识别。首先在大规模的非限制人脸数据上进行网络预训练,然后在目标监控人脸数据上进行微调。其次,网络结构全部采用卷积层,去除全连接层,从而将网络参数量控制在一定范围内。因此,过拟合问题得到很好的解决。此外,多网络的输出在特征层进行融合从而获得更具鉴别性的特征。同时,采用人脸姿态虚拟生成对注册集数据进行增广,从而增强对姿态的鲁棒性。基于以上改进,该算法在监控人脸识别上取得较好的性能,并且在ICB-RW2016竞赛中获得第三名。
 
3   针对跨模态人脸识别中显著的异质性,提出了异质联合贝叶斯度量学习算法。该算法通过对异质人脸数据的注册集和待验证集的数据分布分别建模,能够学习得到非对称的度量,适用于像近红外对可见光的跨模态人脸识别。由于注册集和待验证集的不同分布建模,数据的异质性能够充分参与计算。该算法在多个跨模态人脸识别数据库中超越了现有算法的性能。
 
4  行人再辨识中,行人数据存在较大的类内差异,导致深度学习的训练受到较大影响,限制其最终性能。针对这一问题,提出了温和正样本挖掘算法,在训练的过程中对训练集难度适中的正样本进行挖掘,从而使得网络收敛更稳定,性能得以提升。另外,还提出了训练中的权重约束,有效地防止了网络的过拟合。基于这些改进,该算法在多个行人再辨识公开库中显著地提升了现有算法的性能。

总结而言,本论文主要贡献在于,提出多个基于深度学习的人脸识别和行人再辨识算法,对现有算法在理论
和实际应用上提出创新方法和改进方法,提升了算法性能。
其他摘要
Face and person recognition is the essential part of biometrics and security.
The current technologies often suffer from various variations in face and pedestrian data, such as illuminations, poses, blurs, low-resolution, occlusions etc. As a result, the performance of recognition is limited in practical applications.
 
In this thesis, our work focuses on the major issues of face and person recognition, and specifically is dedicated for the improvement of face recognition algorithms and person re-identification algorithms that based on deep learning methods. The main contributions of this thesis include the following aspects:

1   Based on auto-encoder, we extend the auto-reconstruction objective to reconstructing a sample from another one of which the labels are identical, namely class-encoder, so to provide effective intra-class constraint in the high dimensional feature space. Moreover, we improve the reconstruction from data-level to feature-level, so that the training process is robust to noises and image translations. Through incorporating the supervised information into the training, the deep class-encoder network learns robust and discriminative features for unconstrained face recognition.
 
2   To cope with the issues of heterogeneity and poor scale of dataset in surveillance face recognition, we propose an ensemble multi-task deep network and its training algorithm. This network is pretrained on the large scale dataset of general face, then finetuned on the target surveillance face data. Besides, the network architecture is build in a full convolution fashion, in order to limit the parameter amount. Therefore, the overfitting problem is well solved. Furthermore, the ensemble networks are fused on the feature-level to gain discriminative features; the galleries are augmented by virtual pose synthesis to obtain the robustness to the pose variations. Due to these improvements, our methods achieves promising performance and ranks in the third in ICB-RW2016 competition.
 
3   To address the problem of significant heterogeneity in the cross-modality face recognition, we propose heterogeneous Joint Bayesian algorithm to learn robust asymmetric metric for such as NIR-VIS face recognition. The proposed method models the gallery and probe with different distributions, in order to take into account the difference inter modalities. Therefore, the improved metric is suitable for heterogeneous face recognition tasks with high performance.
 
4   For person re-identification, there exits the issue of large intra-class variations which harm the training of deep network. To deal with the problem, we propose moderate positive mining algorithm. Since the positives with moderate difficulty are mined for training, the network could converge effectively. Besides, we propose weight constraint to alleviate the overfitting. Due to these improvements, the person re-identification performance is significantly elevated.

In summary, in this thesis, we have achieved a great number of improvements for face recognition and person re-identification in both theoretical and practical aspects.

文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/14787
专题毕业生_博士学位论文
作者单位中国科学院自动化研究所
推荐引用方式
GB/T 7714
石海林. 基于深度学习的人脸及行人识别算法研究[D]. 北京. 中国科学院研究生院,2016.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
mythesis-submit.pdf(16603KB)学位论文 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[石海林]的文章
百度学术
百度学术中相似的文章
[石海林]的文章
必应学术
必应学术中相似的文章
[石海林]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。