一种基于卷积神经网络人脸朝向预判的人脸关键点检测方法

发布时间：2018-01-29 19:03

本文关键词： 卷积神经网络人脸关键点检测人脸对齐级联回归器　出处：《浙江大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着科技的发展,近年来关于人脸的应用层出不穷,而人脸关键点检测,往往是各种人脸应用的基础,其重要性日益凸现。人脸关键点检测(或者说人脸对齐)就是给定一张人脸,用某种方法预测出眼睛,鼻子,嘴巴等一些关键点的位置。在深度学习热起来之前,比较好的方法都是基于回归的方法(由若干个回归器级联而成的boosting算法),而这几年在深度学习的飞速发展下,也有学者提出基于神经网络的算法,取得了很好的效果。在本文章中,我们将卷积神经网络与传统的基于回归的的算法结合在一起,提出了一个新的人脸关键点检测算法,即首先用卷积神经网络对人脸朝向进行分类(分为朝左,朝右,和正面),然后不同的朝向各自训练一个基于LBF特征的级联回归器,从而预测人脸关键点位置。此外,我们还进一步拓展与落实上述思路,提出了用标有68个关键点的数据集去预测5个关键点的做法,即利用更多的信息去获得更好的预测结果。在多个数据集上的实验结果表明,利用卷积神经网络进行人脸朝向的分类能够大幅提高人脸关键点位置检测的正确率,本文提出的算法在运行速度和正确率上都有较为出色的表现,同时在利用标有68个关键点的数据去预测5个关键点的实验中,我们的算法表现非常出色,在AFW数据集上取得了最好的效果。
[Abstract]:With the development of science and technology, the application of human face is emerging in recent years, and face key point detection is often the basis of all kinds of face applications. The key point of face detection (or face alignment) is to give a face, using some method to predict the eyes, nose. The location of some key points such as mouth. Before the deep learning becomes hot, the better method is based on regression (boosting algorithm, which is composed of several regressors cascaded). In the past few years, with the rapid development of in-depth learning, some scholars have also proposed a neural network-based algorithm, which has achieved good results. We combine the convolutional neural network with the traditional regression based algorithm, and propose a new face key point detection algorithm. Firstly, we use convolution neural network to classify face orientation (divided into face to left). Facing to the right and facing, then training a cascade regression based on LBF features to predict the position of the key points of the face. In addition, we further expand and implement the above ideas. In this paper, a method of using data sets marked 68 key points to predict 5 key points is put forward, that is to say, more information is used to obtain better prediction results. The experimental results on multiple data sets show that. The classification of face orientation using convolution neural network can greatly improve the accuracy of face key point location detection. The algorithm presented in this paper has a better performance in both the running speed and the correct rate. At the same time, in the experiment of using the data with 68 key points to predict the five key points, our algorithm performs very well and achieves the best effect on the AFW data set.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【相似文献】