摘要
人脸关键点检测是计算机视觉领域的一个重要分支,其检测精度将在很大程度上影响人脸识别和表情分析的结果.提出一种新的解决人脸关键点检测问题的方法,即H-GBDT.H-GBDT是一种基于GBDT决策树和HOG特征的人脸关键点检测算法,该算法是将人脸图像的HOG特征作为GBDT的输入,关键点的真实坐标作为GBDT的输出来训练预测模型,在该过程中每个关键点将分纵坐标和横坐标两次在GBDT中做回归运算,并经过不断的调整GBDT和HOG特征的参数来训练出最佳预测模型.在BioID、LFW、LFPW三种数据集上验证H-GBDT算法的性能.BioID是正脸数据集,实验结果表明H-GDBT在该数据集上的检测效果最佳,其检测误差基本上可控制在2%以内;而LFW和LFPW是自然场景下的数据集,H-GBDT在这两种数据集上的检测误差一般在2%~4%之间.
The detection of facial points is an important topic in the field of computer vision, and its accuracy directly affects the results of face recognition and facial expression analysis. To solve the problem of facial point detection, this paper proposed a new method, which is hereafter referred to as H-GBDT. H-GBDT is an algorithm used for facial points detection based on GBDT and HOG feature. In H-GBDT, the HOG feature is used as the input of GBDT, and the real coordinates of facial points are used as the output of GBDT in order to train the predict model. Each facial point will be divided into vertical and horizontal coordinates to perform regression in GBDT during training, and through the constant adjustment of the parameters of GBDT and HOG to train the best predict model. This paper will verify the performance of H-GBDT on three datasets, which are BioID, LFW and LFPW. BioID is a frontal face dataset. The experimental results show that the detection effect of H-GBDT is the best on BiolD and the detection error is usually below 2%. The detection error of H-GBDT can be controlled between 2% and 4% on the LFW and LFPW datasets, which are datasets in the wild.
作者
张重生
彭国雯
于珂珂
ZHANG Chongsheng;PENG Guowen;YU Keke(School of Computer and Information Engineering, Henan University, Henan Kaifeng 475001, China)
出处
《河南大学学报（自然科学版）》
CAS
2018年第2期214-222,共9页
Journal of Henan University:Natural Science
基金
国家自然科学基金项目(41401466)
