بعض محتويات هذا التطبيق غير متوفرة في الوقت الحالي.
إذا استمرت هذه الحالة ، يرجى الاتصال بنا علىتعليق وإتصال
1. (CN106845327) Training method and device for face alignment model, and face alignment method and device

المكتب : الصين
رقم الطلب: 102015000894091 تاريخ الطلب: 07.12.2015
رقم النشر: 106845327 تاريخ النشر: 13.06.2017
رقم التسليم: 106845327 تاريخ التسليم: 02.07.2019
نوع النشر: B
التصنيف الدولي للبراءات:
G06K 9/00
Description not available in lang ar
المتقدمون: SPREADTRUM COMMUNICATIONS (TIANJIN), INC.
展讯通信(天津)有限公司
المخترعون: PAN BOYANG
潘博阳
CHEN MINJIE
陈敏杰
LIU YANG
刘阳
GUO CHUNLEI
郭春磊
LIN FUHUI
林福辉
الوكلاء: 北京集佳知识产权代理有限公司 11227
北京集佳知识产权代理有限公司 11227
بيانات الأولوية:
العنوان: (EN) Training method and device for face alignment model, and face alignment method and device
(ZH) 人脸对齐模型的训练方法、人脸对齐方法和装置
الملخص: front page image
(EN) The invention discloses a training method and device for a face alignment model, and a face alignment method and device. The training method comprises the following steps that: utilizing each facial landmark of an ith training sample of which the facial landmarkd are calibrated to train a regressor to correspond to the parameter of the ith training sample; through P rounds of training, obtaining the regressor of the ith training sample; according to the regressor corresponding to the ith training sample, calibrating the facial landmark of a jth training sample; repeating a training process until the regressors independently corresponding to N pieces of training samples are obtained; and taking the regressor corresponding to the last training sample as the face alignment model. When p is greater than or equal to 1 or less than or equal to K, a corresponding linear regression method is a global regression method, and when p is greater than or equal to (K+1) or less than or equal to P, the corresponding linear regression method is a partial regression method. The face alignment model obtained by the above training method is high in accuracy.
(ZH) 一种人脸对齐模型的训练方法、人脸检测方法和装置。所述训练方法包括:利用已标定面部特征点的第i个训练样本的各面部特征点训练回归模型对应于第i个训练样本的参数,经P轮的训练,得到所述第i个训练样本的回归模型,根据所述第i个训练样本对应的回归模型标定第j个训练样本的面部特征点,重复训练过程,直至分别获得N个训练样本对应的回归模型;将所获得的最后一个训练样本对应的回归模型作为所述人脸对齐模型。其中,当1≤p≤K时,所述对应的线性回归的方法为全局回归方法,当K+1≤p≤P时,所述对应的线性回归的方法为部分回归方法。应用上述训练方法得到的人脸对齐模型的精度较高。