基于优化的SIFT特征描述子的人脸特征点定位算法的研究

发布时间：2018-07-20 14:27

【摘要】：人脸图像分析是计算机视觉和模式识别领域的前沿课题,包含人脸检测,人脸特征点定位,表情识别,人脸识别等一系列研究内容。其中人脸检测和面部特征点的精确定位是人脸识别,表情识别,姿态估计等应用中的至关重要的步骤,也是扩大基于人脸分析的应用的范围的过程中需要解决的关键问题。基于这一实际需求,本文所做的工作是在背景中进行人脸定位检测,分割出椭圆形人脸感兴趣区域,然后在检测出的人脸区域内进行基于纹理特征和形状约束的五官特征点精确定位。传统的人脸特征点定位算法对脸部特征点的特有特征研究不多,致使特征点的初始化检测效果不好,从而直接影响到后续对特征点点集的形状约束。针对以上问题,这篇文章根据脸部特征点的纹理特征,利用Powell算法学习得到基于SIFT(Scale Invariant Feature Transform)的特征算子最优化的参数;然后提取脸部特征点在最优化的参数下的SIFT特征算子并用于训练基于支持向量回归机的检测器;最后利用优化的检测器在感兴趣区域内进行特征点初始化检测。文章利用LFPW人脸数据库作为训练库进行SIFT特征算子参数的优化,并训练基于SIFT特征的支持向量回归机。最后在BioID数据库上进行测试,通过实验验证了,在存在光线,姿势,表情变换的情况下,基于改进的SIFT特征训练的支持向量回归机对人脸特征点初始化检测的效果相比于基于未改进的SIFT特征的支持向量回归机有了大幅度的提高,另外引入的基于三点一组模型的外形结构约束算法,成功对人脸特征点初始化检测的异常值进行了校正,进一步提高了人脸特征点定位的精确度。
[Abstract]:Face image analysis is an advanced subject in the field of computer vision and pattern recognition, which includes a series of research contents, such as face detection, facial feature point location, facial expression recognition, face recognition and so on. Among them, face detection and accurate location of facial feature points are important steps in the application of face recognition, facial expression recognition and pose estimation, and are also the key problems to be solved in the process of expanding the scope of application based on face analysis. Based on this practical requirement, the work of this paper is to detect the face in the background and segment the region of interest to the elliptical face. Then the facial feature points based on texture feature and shape constraint are accurately located in the detected face region. The traditional facial feature point localization algorithm has little research on the unique features of facial feature points, which leads to the poor initialization and detection effect of feature points, which directly affects the shape constraints of the feature points set. According to the texture features of facial feature points, this paper uses Powell algorithm to obtain the optimized parameters of feature operators based on sift (scale Invariant feature transform). Then, the sift feature operator of facial feature points under the optimized parameters is extracted and used to train the detector based on support vector regression machine. Finally, the optimized detector is used to initialize the feature points in the region of interest. In this paper, the LFPW face database is used as the training database to optimize the parameters of the sift feature operator and to train the support vector regression machine based on the sift feature. And finally we tested it on the BioID database, and it was verified that, in the presence of light, posture, and facial changes, The effect of SVM based on improved sift feature training for initial detection of facial feature points is much higher than that of support vector regression machine based on unimproved sift feature. In addition, the contour structure constraint algorithm based on 3.1 sets of models is introduced, which successfully corrects the outliers of initial detection of facial feature points, and further improves the accuracy of facial feature points location.
【学位授予单位】：天津大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】