【作者】 曾义；

【导师】 申瑞民；

【作者基本信息】 上海交通大学 ， 计算机软件与理论， 2007， 硕士

【摘要】 人脸检测作为计算机视觉领域的一个热点研究问题,已经被研究了数十年了。早期人脸识别成为学界一个主要的研究议题时,尽管人脸检测作为一个无法绕开的议题被提出来,但普遍被认为难度远低于人脸识别。随着研究的深入,人脸检测被认为难度及研究意义与人脸识别相当,并逐渐发展为模式识别与图像处理领域的一个独立研究分支。人脸检测一方面作为人脸识别的前驱工作,有其无可替代的应用价值。另一方面,由于安全监控、自动跟踪等领域的需求,对视频中人脸检测的实时性要求也越来越高。2001年Paul Viola提出的结合Boosting与级联分类的算法使得人脸的实时检测向前迈进了一大步,但是整个冗长的训练过程和检测仍有提高的空间。提高检测分类器的性能对人脸检测实时性的提升有着直接的贡献,而缩短训练过程对探索新的高效检测技术有着十分重大的意义。因此本文在现在文献的基础上实现了一个高准确率高性能的实时人脸检测系统。为了减少训练时间,我们提出了一种基于Boosting的改进训练算法。该算法通过对弱分类器的阈值选择进行一趟处理来降低运算时间复杂度,并根据Boosting训练迭代中只改变样本权值而不更新样本的特点对特征值和排序结果进行缓存来提高训练算法的更多还原

【Abstract】 As a hot issue in the research field of computer vision, face detection has been studied for decades. In the early years, face detection, though was proposed as the first step of face recognition, was viewed as far easier than the latter. Nevertheless, as the study goes deeper, face detection now is considered to be as difficult as, if not more difficult than, face recognition.As the pre-process of face recognition, face detection has its unalterable value. On the other hand, for its application in the field of biometric surveillance and automatic tracking, the real time requirement of face detection is becoming higher. Although the proposal of the combination of Boosting algorithm and cascade classifier by Paul Viola drew a great advance of real time face detection, there are still space for the improvement of classifier training and detection.Therefore an improved training algorithm of AdaBoost is proposed to solve the problem of long period of training in this thesis. According to this improved algorithm, threshold selection of weak classifier is processed in a single pass to reduce time complexity. And computational results of sorted feature values are buffered, cutting down the number of re-calculations significantly.Also a novel cascaded classifier constructing method is proposed更多还原