摘要
针对人脸特征点定位容易受到人脸姿态、表情、光照和遮挡等因素影响的问题,提出了一种基于随机森林回归的特征点定位方法。采用像素差值特征在每个特征点建立随机森林模型,由森林模型回归得到训练样本的估计形状;对训练样本的估计形状与真实形状进行线性最小二乘拟合,得到一个全局优化模型;再利用模型对测试样本特征点位置进行回归估计及形状优化,从而实现了人脸特征点的自动定位。该方法采用相对误差小于0.1的衡量标准,在MATLAB R2009a平台实验,Helen和LFPW数据库上测试,两个数据库的样本定位正确率均超过95%,平均定位速度为9.3 fps,且训练模型仅为5.4 MB。实验结果表明:提出的方法能够很好地适应人脸姿态、表情、光照和遮挡等因素的影响,同时训练模型小,且有效地提高人脸特征点定位的精度和速度。
A method for the facial landmarks location based on the random forest regression is introduced. And the facial landmarks location is easily suffered from the influences of some factors,such as the face poses,expressions,illuminations and partial occlusion,etc. The pixel-difference feature at each landmark is adopted to establish the random forest regression model,and these local models are regressed to predict the shapes of training samples. The estimated shapes and ground truth shapes of the training samples are fit by using the linear least squares method to get an effectively global optimization model. The models above are used to regression estimate on feature points position and shapes optimization for testing sample,then the facial feature points can be automatically located. The method adopts a standard that is relative error 0. 1,conducting experiments in Matlab R2009 a and tests in both Helen database and LFPW database,the location accuracys of the two databases that are all over 95% are obtained,and the average location speed is 9. 3 fps,besides,the train model capacity is only 5. 4 MB. The experimental results show that the introduced method can primely adapt to the influences coming from factors of the face poses,expressions,illuminations and partial occlusion,etc. Simultaneously,whose size of the train model is quite small and effectively improves the accuracy and speed of the facial landmarks location.
出处
《电子测量与仪器学报》
CSCD
北大核心
2016年第5期684-693,共10页
Journal of Electronic Measurement and Instrumentation
基金
国家自然科学基金(61175075
61271382)
国家高技术研究发展计划(863计划)(2012AA112312)
国家科技支撑计划(2015BAF13B01)资助项目
关键词
人脸特征点定位
随机森林
级联回归
像素差值特征
facial landmarks location
random forest
cascaded regression
pixel-difference features
