基于深度学习的人脸特征点定位和人脸识别算法的研究与实现

发布时间：2018-07-10 13:19

本文选题：人脸特征点定位 + 深度学习　；参考：《北京邮电大学》2016年硕士论文

【摘要】：近几年,随着互联网的飞速发展和计算机视觉算法研究的不断深入,基于深度学习的图像处理技术得到了蓬勃的发展。人脸特征点定位和人脸识别作为计算机视觉主要研究方向之一,也得到了很大发展。在这种背景下,本文在前人的研究基础上,实现并改进了基于自适应级联卷积神经网络的特征点定位。本文的研究工作主要有以下几个方面:1.对级联线性回归方法进行了总结与分析,指出级联系统对人脸特征点定位是十分有效的。2.提出了一个用于人脸特征点定位的新的卷积神经网络结构,同时提出了自适应级联卷积神经网络系统,使用高斯模型对级联系统中的前一层网络的输出进行建模,用于选择下一层网络的训练样本,使得级联变的高效。3.基于自适应级联算法,实现了用于人脸脸部五个特征点检测和六十八个特征点的系统。4.基于特征点定位,实现基于卷积神经网络的人脸特征提取,以及人脸识别算法,在标准数据集LFW上达到97.2%的准确率。
[Abstract]:In recent years, with the rapid development of the Internet and the research of computer vision algorithm, image processing technology based on in-depth learning has been booming. As one of the main research directions of computer vision, face feature point location and face recognition have also been greatly developed. In this context, based on the previous research, the feature point localization based on adaptive cascaded convolution neural network is realized and improved. The research work of this paper mainly has the following several aspects: 1. The concatenated linear regression method is summarized and analyzed, and it is pointed out that the cascade system is very effective in locating facial feature points. In this paper, a new convolution neural network structure for face feature point location is proposed, and an adaptive cascaded convolution neural network system is proposed. The output of the previous layer network in the cascade system is modeled by using Gao Si model. The training sample used to select the next layer network makes the cascade variable efficient. 3. Based on the adaptive cascade algorithm, the system. 4. 4 is implemented to detect five feature points and 68 feature points of face. Based on feature point location, face feature extraction based on convolutional neural network and face recognition algorithm are realized. The accuracy rate is 97.2% on LFW standard data set.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41;TP181

【相似文献】