基于人脸图像的年龄估计

CASIA OpenIR > 毕业生 > 硕士学位论文

	基于人脸图像的年龄估计
	刘婷
	2016-05-25
学位类型	工学硕士
中文摘要	年龄作为人体的一种重要生物特征，在安全监控、人机交互、视频检索等领域有着巨大的应用潜力，并且是人脸识别技术的主要瓶颈问题之一。基于人脸图像的年龄估计技术作为一种新兴的生物特征识别技术，目前已经成为计算机视觉领域的一个重要研究课题。本文主要围绕基于传统方式的年龄敏感特征提取、基于卷积神经网络的人脸年龄估计和构建人脸年龄数据库等三个方面开展研究工作。主要研究成果和贡献如下： (1) 提出了判别式人脸描述子网络结构（Discriminant Face Descriptor Network,DFDnet），学习更具有判别力的年龄敏感特征，提高表达能力。由于人的老化过程是随着时间缓慢变化的，同一年龄段的人纹理形状特征相近，例如皱纹、老年斑等，学习年龄敏感特征显得尤为重要。传统特征的提取模式一般都是预先定义的，本文提出的DFDnet算法是数据驱动的，学习判别式特征，使得同一年龄的图像特征差别较小，不同年龄的图像特征差别较大，提取的年龄相关特征可以产生更有判别力的结果。且DFDnet为多层堆叠结构，可以挖掘更深层的信息，从而获得更为精确的预测年龄值。 (2) 提出了基于多区域卷积神经网络（Multi-Region Convolutional Neural Network，MRCNN）的人脸年龄估计算法，实现多个人脸子区域联合学习。与依据整个人脸图像进行表示不同，本文提出局部区域特征表示算法，并进行针对性的分析，使得人脸特征表达更加鲁棒。依据不同人脸局部区域对于年龄估计贡献度的大小，同时考虑到人脸的对称性，共筛选出8组子区域联合用于年龄估计。MRCNN针对这8组局部人脸区域，分别构建8个子网络结构，然后在特征层融合用于年龄估计。MRCNN网络结构有两方面的优势，一是8个子网络可以学习每块局部人脸区域独有的特征，二是最后的全连接层将所有子网络结合在一起，相互补充，可以获得更多的年龄信息。同传统方法相比较，MRCNN不需要人工设计特征，它具有自动学习特征的优点，学到的特征更为有效。实验表明，MRCNN相对于传统的估计方式具有更好的预测性能。 (3) 提出构建人脸年龄数据库，跨越0-100岁所有年龄。建立具有年龄信息的人脸库是人脸年龄估计研究的第一步，现有的人脸年龄数据库年龄分布不均匀，年龄较小或较大的人脸图像很少。没有充分的样本数据，使训练出的年龄估计模型的推广性很差。因此，本文提出构建大年龄跨度的人脸库。首先通过网页分析从互联网上获取具有人脸以及年龄信息的图像，然后利用人脸检测器对这些图像进行人脸检测，提取人脸图像样本。然后通过“机器+人工”的方式对下载的图片进行筛选，以获得尽可能准确的年龄信息。将筛选出来的图像作为训练样本通过“深度学习+人工”的方式逐步筛选扩大数据集。
英文摘要	As one of the most important biologic features, age has tremendous application potential in various areas such as surveillance, human-computer interface and video detection, and it is one of main bottleneck problems of face recognition technology. Age estimation based upon face images, as an emerging biometrics identification technology, has become a hot topic among computer vision areas. This article conducts the researches in three aspects: age-sensitive feature extraction, age estimation based on Convolutional Neural Networks and the construction of human age database. The main works and contributions of this thesis are summarized as follows: (1) For the purpose of learning more discriminative age-sensitive features, we propose a novel architecture named Discriminant Face Descriptor Network (DFDnet), which can improve the ability of age presentation. As aging process is slowly changing over time, images of the same age group have similar texture and shape features, such as wrinkles, age spots. As a result, learning agesensitive features is of high importance. Traditional feature extraction methods are pre-defined, but our proposed DFDnet method is data-driven, which can learn features of great discriminative information. DFDnet method extracts age related features which minimize the differences within images of the same age while maximize the differences among images with different ages. Besides, DFDnet belongs to stacked structure, which can excavate deeper information, and consequently obtain more accurate age prediction results. (2) A novel Multi-Region Convolutional Neural Network (MRCNN) is proposed. Multiple face subregions join together to estimation age. Compare with global face representation, local-face representation is robust against rotation etc. image transformation. Each targeted region is analyzed to explore the contribution degree to age estimation. According to the face geometrical property, we select 8 subregions, and construct 8 sub-network structures respectively, and then fuse at feature-level. The proposed MRCNN has two principle advantages: 8 sub-networks are able to learn the unique age characteristics of the corresponding sub-region and the final fully connected layer packages the eight together to complement age-related information. Unlike existing traditional methods, MRCNN automatically learns effective features instead of manual designed features. Experiments show the effectiveness of MRCNN. (3) An age database is constructed，which spans all age 0-100 years. Constructing face databases containing age information is the first step of age estimation research. The existing age databases distributions are uneven, as the younger and the older people are very few. For the lack of abundance samples, the estimation models have poor generalization. So we propose to build a large age span database. First we obtain face images and their ages from Internet by analyzing the dynamic web. Then, face detector are utilized to detect face. We use the way of "Machine selection plus manual exclusion" to filter the detected image, to make age information of the filtered images as accurate as possible. Treating the filtered image set as the training set, we further expand the data set step by step by "deep learning + manual exclusion" method.
关键词	人脸年龄估计特征提取卷积神经网络人脸年龄数据库
文献类型	学位论文
条目标识符	http://ir.ia.ac.cn/handle/173211/11526
专题	毕业生_硕士学位论文
作者单位	中科院自动化研究所模式识别国家重点实验室
第一作者单位	模式识别国家重点实验室
推荐引用方式 GB/T 7714	刘婷. 基于人脸图像的年龄估计[D]. 北京. 中国科学院大学,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
刘婷+基于人脸图像的年龄估计.pdf（3280KB）	学位论文		限制开放	CC BY-NC-SA