基于肤色模型的人脸检测及特征点定位方法研究

发布时间：2018-09-12 14:27

【摘要】：人脸检测及特征点定位在计算机视觉研究中有着重要地位,在生活中应用也极其广泛。人脸检测是指在输入图像中检测是否存在人脸,如果存在,则标识出人脸区域的位置。特征点定位则是在人脸检测的基础上,更精确地寻找到脸部的特征点位置。人脸检测及特征点定位是很多人脸相关应用如表情识别、姿态估计、人脸动画合成中的关键步骤,这两个步骤的性能对其后续人脸相关应用有重要影响。传统的人脸检测及特征点定位方法在训练阶段需要进行复杂的特征提取,且鲁棒性不佳。本论文致力于简化特征提取环节,并构造新的特征点定位优化方法,进一步提高检测和定位的准确性和速度。本文对人脸检测及特征点定位方法进行了深入研究。首先,将肤色模型和用深度学习方法训练的人脸分类器结合用于人脸的检测,在人脸检测的基础上,再利用回归网络对检测结果不精确的人脸区域做回归处理,以获得更精准的人脸检测定位。接下来,根据回归后的人脸检测结果进行特征点定位。特征点定位是采用随机森林方法,通过对人脸特征点建立全局约束模型进行整体优化,利用级联回归结构迭代获得人脸特征点的精确位置。论文主要研究内容如下:1、而本文中通过深度学习方法设计并训练一个人脸分类器网络,将其与肤色模型相结合,能够更有效地检测出复杂场景下的人脸,且避免了显式的特征设计和提取环节。2、针对前一步骤人脸检测环节中可能出现的检测精度不高的情况,设计了一个回归网络,利用回归网络对检测结果进行回归校正以获得更精确的检测定位。3、特征点定位初始阶段将根据人脸检测结果给特征点赋予初始坐标,因此,一个精确的人脸检测结果对于提升特征点定位的速度和精度有重要作用。本文首先训练随机森林模型实现对特征点定位重要特征的筛选,之后对人脸特征点建立全局约束模型,用最小二乘方法对全局模型参数进行整体优化,最后利用级联回归结构进行迭代获得人脸特征点的精准定位。实验结果表明,改进的人脸检测及特征点定位系统能够有效提高复杂环境下人脸检测及特征点定位的性能,在保证鲁棒性及较高定位精度的前提下,还拥有接近实时的较高检测及定位速度。
[Abstract]:Face detection and feature location play an important role in computer vision research and are widely used in daily life. Face detection is to detect the presence of a face in an input image and, if so, to identify the location of the face region. On the basis of face detection, feature point location is more accurately located. Face detection and feature location are the key steps in many human face related applications such as facial expression recognition, pose estimation and face animation synthesis. The performance of these two steps plays an important role in the subsequent face related applications. The traditional face detection and feature point localization methods need complex feature extraction in the training stage, and the robustness is not good. This paper is devoted to simplify feature extraction and construct a new method of feature location optimization to further improve the accuracy and speed of detection and location. In this paper, the methods of face detection and feature point location are studied. Firstly, the skin color model and the face classifier trained by the depth learning method are combined for face detection. Then, based on the face detection, the regression network is used to deal with the face region with inaccurate detection results. In order to obtain more accurate face detection location. Then, the feature points are located according to the result of face detection after regression. The feature point location is based on the stochastic forest method. The global constraint model is established for the face feature points and the exact location of the face feature points is obtained by cascading regression structure iterations. The main contents of this paper are as follows: 1. In this paper, a human face classifier network is designed and trained by the deep learning method, which is combined with the skin color model to detect the human face in the complex scene more effectively. And avoid the explicit feature design and extraction link. 2. Aiming at the situation that the detection accuracy may not be high in the previous step face detection link, a regression network is designed. Regression network is used to correct the detection results to obtain more accurate detection location. The initial phase of feature point location will be assigned to the initial coordinates of feature points according to the face detection results, so, An accurate face detection result plays an important role in improving the speed and accuracy of feature point location. In this paper, we first train the stochastic forest model to select important features for feature point location, and then establish a global constraint model for facial feature points, and optimize the global model parameters by using the least square method. Finally, a cascade regression structure is used to iterate to obtain accurate location of face feature points. The experimental results show that the improved face detection and feature point location system can effectively improve the performance of face detection and feature point location in complex environments, while ensuring robustness and high positioning accuracy. It also has high detection and positioning speed near real time.
【学位授予单位】：重庆理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】