基于深度学习自动年龄估计技术研究

发布时间：2018-03-22 13:17

本文选题：深度学习　切入点：年龄估计　出处：《华侨大学》2017年硕士论文　论文类型：学位论文

【摘要】：人脸年龄估计已经成为机器视觉领域的一个研究热点。国内外许多学者都对其进行大量的研究,但是由于年龄变化是一个极其复杂的过程,它受到生活环境、基因遗传等因素影响。通过人脸图像来预测年龄值依然是一个非常具有挑战的课题。研究人脸年龄估计可分为年龄特征的提取和年龄的估计。根据目前的研究状况,本文从提取描述能力更强的年龄特征和使用回归模型来预测年龄值上进行深入研究。全文的主要工作概括如下:1)提出一个新的特征学习方法,称为PCANet特征学习模型。PCANet是一种简单的深度学习模型,它通过PCA算法学习出卷积层的卷积核,然后使用该模型提取人脸图像的年龄特征。深度学习结构已经被证明具有非常出色的学习能力,它能够捕捉到人脸年龄外貌形态变化的特征和学习到具有强大描述能力的年龄特征。因此,PCANet特征学习模型可以在人脸图像年龄估计上取得更好的结果。2)提出一种多层深度学习PCA网络模型特征的表达方法。深度网络是以逐层的方式进行特征表示,网络的低层能够学习到边缘特征,高层能够学习到抽象特征。所以多层PCA网络模型特征的表达是融合底层特征和高层特征作为最终的年龄特征值,并且利用高低层特征的表达能力来提取到更加丰富的年龄特征信息。3)提出一种深度迁移训练模型的年龄估计方法。针对深度学习模型在小数据集条件下训练可能会产生过拟合现象而提出,它能有效的解决过拟合问题。该方法是先用ImageNet图像集预训练网络模型,再将学习ImageNet图像集得到的权重作为网络的初始权重。然后再用表观人脸年龄数据集微调该通用网络模型使其具备学习年龄特征的能力,从而提高人脸年龄预测的效果。
[Abstract]:Face age estimation has become a research hotspot in the field of machine vision. Many scholars at home and abroad have done a lot of research on it, but because age change is an extremely complex process, it is subjected to living environment. It is still a challenging task to predict age value by face image. The study of face age estimation can be divided into age feature extraction and age estimation. In this paper, a new feature learning method is proposed by extracting more descriptive age characteristics and using regression model to predict the age value. The main work of this paper is summarized as follows: 1). Called PCANet feature learning model. PCANet is a simple depth learning model. It uses PCA algorithm to learn the convolution kernel of convolution layer. Then the model is used to extract the age characteristics of the face image. The depth learning structure has been proved to have excellent learning ability. It can capture the feature of facial age appearance change and learn the age feature with strong descriptive ability. Therefore, PCANet feature learning model can obtain better results in age estimation of face image. 2) A new method is proposed. The method of expressing the features of multi-layer depth learning PCA network model. The lower level of the network can learn the edge feature, the higher level can learn the abstract feature, so the expression of the multi-layer PCA network model features is the fusion of the bottom feature and the high-level feature as the final age characteristic value. And using the expression ability of high and low level features to extract more abundant age feature information. 3) A method of estimating the age of depth transfer training model is proposed. The training possibility of depth learning model under the condition of small data set is discussed. It is proposed that the phenomenon of over-fitting will occur. It can effectively solve the problem of over-fitting. This method uses ImageNet image set to pretrain the network model. Then the weight obtained from the learning ImageNet image set is taken as the initial weight of the network, and then the apparent face age data set is used to fine-tune the general network model so that it has the ability to learn the age characteristics, thus improving the effect of face age prediction.
【学位授予单位】：华侨大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】