face-expression-detect

github代码:https://github.com/adithyaselv/face-expression-detect

原理

人脸表情分类,将人脸表情分为7类,

emotions={ 1:"Anger", 2:"Contempt", 3:"Disgust", 4:"Fear", 5:"Happy", 6:"Sadness", 7:"Surprise"}

分类原理为,首先采用dlib检测人脸的68个特征点,根据特征点提取人脸特征向量,之后采用pca对特征向量进行降维最后采用SVM对输入特征向量进行分类.

根据特征点提取人脸特征向量,

features=generateFeatures(landmarks)

得到一个26565维的特征向量后,采用pca进行降维,得到一个75维的特征向量

print "Performing PCA Transform......."
pca_features=pca.transform(features)

之后将75维的特征向量输入SVM进行分类(这里分为)

根据人脸23个的特征点,将这些点每4个组合成一个子序列,这样得到8855中组合.

keyPoints = [18, 22, 23, 27, 37, 40, 43, 46, 28, 32, 34,
             36, 5, 9, 13, 49, 55, 52, 58,61,63,65,67]
combinations = itertools.combinations(keyPoints, 4)#将keyPonints中的所有元素组合成四个元素的子序列,这样有C23^4=8855种组合

对于每个子序列,按照三种方式重新排序,这样得到$8855\times3=26565$种特征点组合.

pointIndices1 = []
pointIndices2 = []
pointIndices3 = []
pointIndices4 = []#26565

for combination in combinations:
    pointIndices1.append(combination[0])
    pointIndices2.append(combination[1])
    pointIndices3.append(combination[2])
    pointIndices4.append(combination[3])

    pointIndices1.append(combination[0])
    pointIndices2.append(combination[2])
    pointIndices3.append(combination[1])
    pointIndices4.append(combination[3])

    pointIndices1.append(combination[0])
    pointIndices2.append(combination[3])
    pointIndices3.append(combination[1])
    pointIndices4.append(combination[2])

根据得到的4组坐标点索引组合,分别为pointIndices1, pointIndices2, pointIndices3, pointIndices4,每个组合长度为26565.根据坐标点索引,计算特征点的位置坐标x,y值. 最后根据每两个坐标点之间的欧式距离之比计算特征向量:

Features = numpy.zeros((1, len(pointIndices1)))
ratios = [];

for i in range(0, len(pointIndices1)):
    x1 = landmarkCoordinates[2*(pointIndices1[i]-1)]
    y1 = landmarkCoordinates[2*pointIndices1[i] - 1]
    x2 = landmarkCoordinates[2*(pointIndices2[i]-1)]
    y2 = landmarkCoordinates[2*pointIndices2[i] - 1]

    x3 = landmarkCoordinates[2*(pointIndices3[i]-1)]
    y3 = landmarkCoordinates[2*pointIndices3[i] - 1]
    x4 = landmarkCoordinates[2*(pointIndices4[i]-1)]
    y4 = landmarkCoordinates[2*pointIndices4[i] - 1]

    points = [x1, y1, x2, y2, x3, y3, x4, y4]
    ratios.append(findRatio(points))#计算x1,y1与x2,y2的欧式距离d1,计算x3,y3与x4,y4的欧式距离d2,ration=d1/d2

Features = numpy.asarray(ratios)#26565
return Features

计算特征值:

def findRatio(points):

    ''' Function to find ratios '''

    x=[0]*4
    y=[0]*4
    for i in range(4):
        x[i]=points[(2*i)] 
        y[i]=points[(2*i)+1]

    dist1 = math.sqrt((x[0]-x[1])**2 + (y[0]-y[1])**2)
    dist2 = math.sqrt((x[2]-x[3])**2 + (y[2]-y[3])**2)
    ratio = dist1/dist2

    return ratio

模型训练

由上面分析可知,我们需要训练的有PCA, SVM两个模型.

训练数据,下载包括四个文件夹,

/media/qinghua/文档/data/data_image/ck/cohn-kanade-images,人脸表情图像.

/media/qinghua/文档/data/data_image/ck/Landmarks,为每张图像的特征点坐标x,y.

/media/qinghua/文档/data/data_image/ck/Emotion,保存了每张图像对应的表情标签,分别为1-7范围内的值,

保存为文本文件,文件名为图像名.