【作者】 贺灏；

【导师】 贺建飚；

【作者基本信息】 中南大学 ， 计算机应用技术， 2009， 硕士

【摘要】 人脸检测是人脸分析的首要环节,其处理的问题是确认图像中是否存在人脸,如果存在则对人脸进行定位。人脸检测的应用领域相当广泛,是实现机器智能化的重要步骤之一。AdaBoost算法是2001年提出的一种快速人脸检测算法,是人脸检测领域里里程碑式的进步,AdaBoost算法是一种可以将弱学习转化为强学习的方法,从理论上讲,只要有足够多的样本,足够多的特征,训练足够充分,AdaBoost训练出来的分类器的错误率可以无限趋于零。但是,正因为如此,当样本数目比较多,特征数目也很多时,AdaBoost训练算法存在训练时间太长的问题。同时,在检测人脸过程中,由于大多数的检测算法采用穷举方式,当原始图片过大时,也存在检测时间长的问题。本文阐述了对AdaBoost算法的三点改进方法。第一,通过分解二维特征矩阵,将并行计算引入AdaBoost训练算法当中,训练速度在多机环境下可以得到显著提高;第二,将被检测区域划分成多块,用并行检测代替传统的串行检测,这种改进可以在多核处理器上显著提高人脸检测的速度;第三,改进原有的串行人脸检测算法中的移动步长策略,用前一次检测中通过的强分类器数目来动态决定下一次的移动步长,去除许多明显没有意义的检测,从而提高串行人脸检测算法的检测速度。更多还原

【Abstract】 Face detection is the first phase of face anlysis, the problem refer to face detection is to determinate whether there are human faces in the image, if so, then locate the human faces in the image. Face detection can be used in many fields, and it is one of the most important steps to implement machine intelligence.AdaBoost algorithm is a fast face detection algorithm presented in 2001, it is a mile-stone in the field of object detection. AdaBoost theory can transform weak learning to strong learning, theoretically, if there are enough samples, enough features, and the training is absolutely adequate, the error rate of the classifiers that generated by AdaBoost algorithm is unlimited near zero. But, as the number of samples increases and the number of features increases, the training time of AdaBoost algorithm becomes incredible long. Also, because the standard detection algorithm detects the object by searching the entire image one subwindow by one subwindow. When the image is too big, the detection time is also too long, so it can’t be used in a real-time environment.This paper describes three improvements on the AdaBoost algorithm. Firstly, through the decomposition of the two dimensional feature matrix and the using of parallel computing in AdaBoost training algorithm, the training speed will be increased significantly in multi-machine training. Secondly, by the replacement of traditional serial detection with parallel detection, the detection speed will be increased remarkably when multi-processor machine is used for detection. Thirdly, by using a dynamic step strategy to replace the constant step strategy, the number of windows to be detected will be decreased observably, so the detection speed will be increased significantly.更多还原