基于可变形模型的人脸检测算法研究与应用

发布时间：2018-06-05 06:10

本文选题：人脸检测 + 可变形模型　；参考：《南京理工大学》2017年硕士论文

【摘要】：人脸面部图像分析技术是计算机视觉热点问题之一,而人脸检测是诸多面部分析技术的前提。就目前而言,正面人脸检测技术已趋近成熟,研究成果广泛应用于人脸识别、生物信息安全、虚拟现实等领域,而多视角、遮挡、表情夸张等自然状态下的人脸快速检测技术还相对薄弱。本文在分析和总结基于传统可变形模型的人脸检测方法优缺点的基础上,引入人脸对齐算法,提出一种二级级联人脸快速检测框架,并分别将支持向量机和级联提升算法应用到检测框架,评测结果表明检测器能加快人脸检测速度,提高检测精度,有效降低误检率。本文主要的创新工作如下:(1)结合传统可变形模型的优点,引入人脸对齐算法,提出二级级联人脸快速检测框架。框架包含两个层级,第一层级检测器快速稳定的提供人脸候选框;第二层级检测器通过学习以人脸特征点为中心的判别信息强的局部区域特征提高精度,过滤第一层级得到的人脸候选框,去除非人脸子窗口。框架的三大优势:级联方式提高检测速度;无需对多视角人脸建立相应模型;训练样本易搜集。(2)本文将支持向量机应用于二级级联检测框架,提出基于二级级联支持向量机的人脸快速检测算法。在第一层级中,设计了基于图像梯度的人脸稀疏特征;在第二层级中,算法依据人脸特征点分别提取SIFT和SURF特征,实验对比分析两种特征下的检测精度和速度,并指出可引入高性能人脸检测算法改善检测器性能。训练过程中,本文设计了有效的难例挖掘方法加速收敛。算法采用C++开发,并采用了多个方法加速检测过程,与其他高性能算法评测对比,结果表明二级级联支持向量机人脸快速检测算法能流畅运行于VGA视频流当中,能处理姿态变化的人脸,且有效降低误检,为无约束条件下的人脸检测提供一种有效解决方案。(3)针对传统基于级联Boosting的人脸检测算法误检率稍高,人脸检测和对齐联合训练时正样本难搜集的问题,本文提出基于人脸对齐感知的级联Boosting人脸检测算法。该算法第一层级通过模拟近视患者在模糊视野下筛选人脸的能力去除大量背景子窗口,第二层级通过特征点周围的菱形区域提取局部特征,并尝试解决人脸对齐失败的情况。经过评测集测试,实验结果显示,与传统算法相比该算法误检率低,且能解决人脸姿态变化、轻微遮挡等问题,达到了较高的精度。
[Abstract]:Face image analysis is one of the hot issues in computer vision, and face detection is the premise of many face analysis techniques. At present, the frontal face detection technology has come to maturity, and the research results are widely used in face recognition, biological information security, virtual reality and other fields, but multi-angle, occlusion, etc. Expression exaggeration and other natural state of face detection technology is relatively weak. On the basis of analyzing and summarizing the advantages and disadvantages of face detection methods based on traditional deformable model, this paper introduces face alignment algorithm, and proposes a two-stage cascaded fast face detection framework. Support vector machine (SVM) and cascaded lifting algorithm are applied to the detection framework respectively. The evaluation results show that the detector can accelerate the speed of face detection, improve the detection accuracy and effectively reduce the false detection rate. The main innovations of this paper are as follows: (1) combining the advantages of the traditional deformable model and introducing the face alignment algorithm, a two-stage cascaded fast face detection framework is proposed. The framework consists of two levels: the first layer detector provides the face candidate frame quickly and stably; the second level detector improves the accuracy by learning the local region features with strong discriminant information centered on the face feature points. Filter the first level to get the face candidate, go unless the face subwindow. Three advantages of the framework: cascading method to improve detection speed; no need to build a corresponding model of multi-view face; training samples easy to collect. 2) this paper applies support vector machine to two-level cascade detection framework. A fast face detection algorithm based on two-stage cascade support vector machine is proposed. In the first level, a face sparse feature based on image gradient is designed. In the second level, the SIFT and SURF features are extracted according to the face feature points, and the detection accuracy and speed under the two features are compared and analyzed. It is pointed out that the high performance face detection algorithm can be introduced to improve the performance of the detector. In the process of training, this paper designs an effective mining method to accelerate convergence. The algorithm is developed with C, and several methods are used to accelerate the detection process. Compared with other high performance algorithms, the results show that the two-stage cascaded support vector machine (SVM) fast face detection algorithm can run smoothly in the VGA video stream. It can deal with human face with changing pose and reduce false detection effectively. It provides an effective solution for face detection under unconstrained conditions. (3) the traditional face detection algorithm based on cascaded Boosting has a slightly higher false detection rate. It is difficult to collect positive samples in face detection and alignment joint training. A cascade Boosting face detection algorithm based on face alignment perception is proposed in this paper. The first level of the algorithm removes a large number of background subwindows by simulating myopia patients' ability to screen faces in fuzzy field. The second level extracts local features by diamond region around feature points and tries to solve the problem of face alignment failure. The experimental results show that compared with the traditional algorithm, the algorithm has lower false detection rate and can solve the problems of face pose change and slight occlusion, and achieve a higher accuracy.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】