非约束环境下基于小样本的人脸特征精确定位  

Precise Facial Feature Localization under Non-Restraint Environment with Limited Training Images

作　　者：陈莹[1,2] 张龙媛[1] 

机构地区：[1]江南大学物联网工程学院轻工过程先进控制教育部重点实验室,无锡214122 [2]上海交通大学系统控制与信息处理教育部重点实验室,上海200240

出　　处：《模式识别与人工智能》2013年第6期577-583,共7页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金项目(No.61104213);江苏省自然科学基金项目(No.BK2011146)资助

摘　　要：针对非约束环境下的人脸特征定位问题,在概率框架下提出一种基于小样本的精确定位策略.通过对比分析,提取人脸主要特征的颜色和灰度信息及人脸特征之间的几何约束信息,利用混合高斯模型分别对其进行概率建模.之后建立定位融合策略,不仅考虑每种人脸特征的概率分布,还考虑其周围元素的概率分布特性,及各元素之间的几何约束.实验结果表明,该方法能在少量训练样本图像且样本个体较为单一的条件下,实现人脸主要特征的精确定位,且定位精度高于现有方法.After analyzing the limitation of current methods, a precise localization strategy with limited training data is proposed in a probability framework. Texture and geometry information of facial elements are extracted as model features after comparison analysis with other traditional descriptors. Gaussian mixture model is used for the probability modeling, which describes the distribution of each model features extracted from different facial conditions well. Then, a series of fusion strategies are designed for the facial features localization, which considers the probability distribution of each facial feature, the distribution characters of their surrounding elements and their geometry constraints. The experimental resuhs show that the proposed method can realize precise localization for the facial features with limited training sample images which belong to a single subject, and it outperforms other methods in localization accuracy.

关 键 词：非约束环境 小样本 人脸特征定位 高斯混合模型 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]