臉型匹配

方法1，基於曲率，計算曲率組成的特徵向量的餘弦距離

def cos_sim(vector_a, vector_b):
    """
    計算兩個向量之間的餘弦相似度
    :param vector_a: 向量 a 
    :param vector_b: 向量 b
    :return: sim
    """
    vector_a = np.mat(vector_a)
    vector_b = np.mat(vector_b)
    num = float(vector_a * vector_b.T)
    denom = np.linalg.norm(vector_a) * np.linalg.norm(vector_b)
    cos = num / denom
    sim = 0.5 + 0.5 * cos
    return sim


def compute_curvature_cov(input_numpy):
    #曲率越大，越彎曲
    #       |y''|
    #k= -----------
    #         3/2
    #   (1+y'2)
    #y’(i) = (y(i+1)-y(i))/h
    #y’’(i) = (y(i+1)+y(i-1)-2*y(i))/h^2

    input_numpy_yijie = (input_numpy[1:,1]-input_numpy[:-1,1])/(input_numpy[1:,0]-input_numpy[:-1,0])
    input_numpy_yijie = input_numpy_yijie[:-1]
    input_numpy_erjie = (input_numpy[2:,1]+input_numpy[:-2,1]-2*input_numpy[1:-1,1])/((input_numpy[2:,0]-input_numpy[:-2,0])*(input_numpy[2:,0]-input_numpy[:-2,0]))
    K = np.abs(input_numpy_erjie) /np.power((np.ones_like(input_numpy_yijie)+input_numpy_yijie*input_numpy_yijie),3.0/2.0)

    return K

方法2，基於人臉關鍵點座標的L1距離，

方法3，基於對比的兩個人臉的關鍵點組成區域的交集佔整個圖片的面積，

def area_sim(keypoints_Xs, keypoints_Xt):
    keypoints_Xt = keypoints_Xt.reshape(-1,2)
    process_num = 17
    keypoints_Xs = keypoints_Xs[:process_num]
    keypoints_Xt = keypoints_Xt[:process_num]

    scale = 256
    mask = np.zeros((scale,scale),np.uint8)
    for i in range(process_num-1):
        triangle = np.array([[[keypoints_Xs[i,0],keypoints_Xs[i,1]], [keypoints_Xs[i+1,0],keypoints_Xs[i+1,1]], [keypoints_Xt[i,0],keypoints_Xt[i,1]]]], dtype = np.int32)
        cv2.fillPoly(mask, triangle, 1)
        triangle = np.array([[[keypoints_Xs[i,0],keypoints_Xs[i,1]], [keypoints_Xs[i+1,0],keypoints_Xs[i+1,1]], [keypoints_Xt[i+1,0],keypoints_Xt[i+1,1]]]], dtype = np.int32)
        cv2.fillPoly(mask, triangle, 1)
        triangle = np.array([[[keypoints_Xt[i,0],keypoints_Xt[i,1]], [keypoints_Xt[i+1,0],keypoints_Xt[i+1,1]], [keypoints_Xs[i,0],keypoints_Xs[i,1]] ]], dtype = np.int32)
        cv2.fillPoly(mask, triangle, 1)
        triangle = np.array([[[keypoints_Xt[i,0],keypoints_Xt[i,1]], [keypoints_Xt[i+1,0],keypoints_Xt[i+1,1]], [keypoints_Xs[i+1,0],keypoints_Xs[i+1,1]] ]], dtype = np.int32)
        cv2.fillPoly(mask, triangle, 1)

    dis = 1- np.sum(mask)/(scale*scale)
    return dis,mask

方法4，對上面的3種方法進行融合，

def total_method(keypoints_Xs, keypoints_Xt):
    #(68, 2) (54,)
    #一共68個關鍵點，取前17個人臉邊緣輪廓點使用
    keypoints_Xt = keypoints_Xt.reshape(-1,2)
    process_num = 17

    #method 1，曲率
    qulv_Xs = compute_curvature_cov(keypoints_Xs[:process_num])
    qulv_Xt = compute_curvature_cov(keypoints_Xt[:process_num])

    weights_matrix = np.ones_like(qulv_Xt, np.float32)
    #weights_matrix[3:12] =2.0
    dis1 = cos_sim(qulv_Xs*weights_matrix , qulv_Xt*weights_matrix)


    #method 2，座標L1
    scale  =256
    keypoints_Xs = keypoints_Xs/scale
    keypoints_Xt = keypoints_Xt/scale

    dis2 = 1 - np.mean(np.abs(keypoints_Xs[:process_num] - keypoints_Xt[:process_num]))#cos_sim(keypoints_Xs[:process_num].reshape(-1,1).squeeze(), keypoints_Xt[:process_num].reshape(-1,1).squeeze())

    #method 3，面積
    dis3 ,mask= area_sim(keypoints_Xs*scale, keypoints_Xt*scale )

    dis = 0.1*dis1+ 0.1*dis2 +0.8*dis3



    return dis

調用，

def compare(points_Xs, face_npy):
    #輸入：68個關鍵點numpy,68個人臉數據庫,注意所有人臉都需要首先進行對齊操作，然後歸一化到256*256
    #輸出：最相似人臉輪廓的索引，人臉輪廓的相似度值，人臉輪廓的差異mask list
    indexs = None
    dis_ls = list()
    mask_list = []
    for j in range(len(face_npy)):  # len(face_npy)
        distz, mask = total_method(points_Xs, face_npy[j])
        dis_ls.append(distz)
        mask_list.append(mask)
    dis_ls_arr = np.array(dis_ls)
    print(dis_ls)
    euc_indexs = np.argsort(-1*dis_ls_arr)
    return euc_indexs, dis_ls ,mask_list

 

效果，

融合>面積>L1距離>曲率