【作者】 吴桂林；

【导师】 周敬利；

【作者基本信息】 华中科技大学 ， 计算机系统结构， 2004， 硕士

【摘要】 人脸检测问题最初作为人脸识别系统的定位环节被提出,近年来由于其在安全访问控制、视觉监测、基于内容的检索和新一代人机界面等领域的应用价值,开始作为一个独立的课题受到研究者的普遍重视。对于人脸这类复杂、难以显式描述的检测模式,基于神经网络的方法具有独特的优势,它把人脸模式的统计特性隐含在神经网络的结构和参数之中,通过对大量样本的训练,来完成检测任务。通过分析研究神经网络的原理和人脸器官的特征,设计了一个输入层节点数为625、输出层节点数为2、隐层节点数为20的三层误差反向传播网络;同时收集了大量不同类型的人脸样本,并对初始人脸样本集中的部分图像进行了一些变换,以提高网络的适应能力;根据误差反向传播网络的特点,对目前普遍采用的非人脸样本收集的自举方法做了一些改进,可以有效地收集到更多的具有代表性的非人脸样本,以进一步提高网络的适应能力。而且在训练过程中引入了学习率这个参数,有效地解决了误差反向传播算法的步长问题。在进行人脸检测之前,对待检测的图像作预处理,以提高检测的准确性。并在窗口扫描过程中采用了基于金字塔的子采样方法来解决图像中人脸的大小和位置问题。实验结果表明,基于误差反向传播神经网络的人脸检测方法可以有效地运用于多人脸、不同大小、不同位置、不同方向、不同面部表情和不同光照条件等情况,同时取得了较高的正确检测率和较低的错误报警率。而且通过适当地增加训练样本的数量和类型,可进一步提高检测性能。更多还原

【Abstract】 In the recent years, the subject of face detection which was originally put forward as location part has been placed a lot of importance as an independent subject by researchers due to the practical value in the fields of safe access control, vision inspection, content-based retrieval and new generation human-machine interface. Neural Network Algorithm which hides the statistical character of face pattern in its structure and parameter has shown its special advantage in the detection pattern which is complicated and difficult to describe such as faces.Three-lay Error Back Propagation Network was designed with 625 nodes in the input lay, 2 nodes in the output lay and 20 nodes in the hidden lay based on the research on the principle of Neural Network and the character of faces and has been improved in the applicability by the training of a large amount of various face examples and non-faced examples which are more representative after the betterment of collecting method. Now, the step problem of Error Back Propagation Algorithm has been effectively solved by the use of learning rate in the training.Before face detection, the veracity of detection is increased by pretreatment of the detection images and the detection faces with different size and position has been solved by the method of pyramid sub-sampling in the process of window scan.The result of experiment shows that the face detection based on Error Back Propagation Neural Network has achieved a high precision detection rate and a low error alarming rate and can effectively detect the images with multi-face, various size, different orientation and pose, various facial expression under various lighting conditions and its detection capability can be improved by increase the amount and types of training examples.更多还原