【摘要】 针对人脸朝向分类这一问题,使用BP神经网络进行判别分析是一个较为成熟的方案,在此基础上,提出了一种新的特征值提取方法。首先探测人脸图像边界并将其转化成二值化的0-1矩阵,分割取出图像中眼睛部分对应的矩阵数据;考虑到人脸图像的特殊性,即头部鬓角的信息数据可能造成干扰,删减相应的矩阵信息;接着进行特征值的提取,取出矩阵中为1的元素分布的"离散程度"和分布位置的平均值形成二维向量;最终以该二维向量为神经网络的输入,5种人脸朝向分类为神经网络的输出,正确识别率可以达到100%。这样的特征值提取方式使特征值具有实际意义,相比于PCA特征值提取法更易理解;无需求出人眼的具体位置,相比于求人眼位置的几何方法更加简洁。更多还原

【Abstract】 BP neural network is a relatively mature solution to face orientation discrimination of a single image.Put forward a new eigenvalue extraction method on the basis of previous studies.Firstly,detect the boundary of the image and make it binary to produce a 0-1 matrix corresponding to pixels.Then,obtain the data related to the surrounding areas of eyes by splitting the matrix.Considering the specialty of face image,that is,the interference from the data on side bums,split the matrix again.Secondly,extract two types of eigenvalue.One is the distribution discrete degree of elements whose value is 1;the other is their average distribution position.Finally,construct the neural network by taking a 2-di-mention vector of the two eigenvalues as input and 5 types of face orientation as output.The correct recognition rate of the network is 100%.Comparing to eigenvalue extraction by PCA,this way is more comprehensible because eigenvalue has more significance in practice.There is no need to find out eye position,so this way is more succinct than geometric methods.更多还原