def test_fer2013_data():
train_image, train_label, test_image, test_label = fdata.load_kaggle_face_data()
for i, mat in enumerate(train_image):
cv2.imshow("mat", mat)
cv2.waitKey(0)
landmark = fapi.get_feature_points_fromimage(mat)
if landmark is None:
print "人脸识别出错"
else:
print "调用face++ api成功"
add_landmark(landmark)
'''
这部分查看一下人脸的特征变形，这里需要计算人脸剖分的方案，需要覆盖全图
'''
point_dict = from_lanmark_to_points(landmark)
triangles = trianglulation(point_dict)
# draw_triangle(mat, triangles, point_dict)
mask = draw_mask(triangles, point_dict, mat.shape)
img = np.array(mask * mat, dtype=np.uint8)
img = cv2.equalizeHist(img)
cv2.imshow("mask", img)
cv2.imshow("mat", mat)
cv2.waitKey(0)
def test_face_trianglulation(mat):
mat = cv2.resize(mat, (400, 400))
landmark = fapi.get_feature_points_fromimage(mat)
# 这里为了能够更好的提取人脸，手动添加两个特征点
if landmark is None:
print "人脸识别出错"
else:
print "调用face++ api成功"
add_landmark(landmark)
add_border_point(landmark, mat.shape)
'''
这部分查看一下人脸的特征变形，这里需要计算人脸剖分的方案，需要覆盖全图
'''
point_dict = from_lanmark_to_points(landmark)
triangles = trianglulation(point_dict)
draw_triangle(mat, triangles, point_dict)
mask = draw_mask(triangles, point_dict, mat.shape)
img = np.array(mask * mat, dtype=np.uint8)
img[0] = cv2.equalizeHist(img[0])
img[1] = cv2.equalizeHist(img[1])
img[2] = cv2.equalizeHist(img[2])
cv2.imshow("mat", mat)
cv2.waitKey(0)
def feature_flow_extract():
"""
人脸区域光流提取
:return:
"""
p = "../dataset/datashow/face_aligen/EP02_01f/"
pf = "../dataset/datashow/flow_aligen/EP02_01f/"
file_name_list = os.listdir(p)
dir = p
for file in file_name_list:
if file[-3:] != "jpg":
continue
img = cv2.imread(dir + file)
landmarks = fd.get_feature_points_fromimage(img)
points = fd.from_lanmark_to_points(landmarks)
fd.supply_landmark(points)
regions = fd.get_region_by_landmark()
points_list = []
mask = np.zeros(img.shape[0:2])
for i in regions:
tmp_point = []
for mark_name in i:
tmp_point.append(points[mark_name])
points_list.append(tmp_point)
fd.draw_region(img, tmp_point)
fd.draw_region_mask(mask, tmp_point)
# 光流
U, V = ft.flow_reader(pf + "reg_flow" + file[3:-4] + ".xml")
flow = np.zeros((U.shape[0], U.shape[1], 2))
value = (U * np.sqrt(1 / 2.0) + V * np.sqrt(1 / 2.0)) * mask
value = cv2.normalize(value, None, 0, 1, cv2.NORM_MINMAX)
plt.imshow(value * mask, cmap="gray")
plt.show()
for i in range(U.shape[0]):
for j in range(U.shape[1]):
if mask[i][j] == 1:
flow[i][j] = (U[i][j], V[i][j])
opt = of.OpticalFlow()
clolc = opt.visual_flow(flow)
cv2.imshow("a", clolc)
cv2.waitKey(0)
cv2.imshow("a", img)
cv2.waitKey(0)
print points_list
def get_masked_face(mat, shape):
landmark = fapi.get_feature_points_fromimage(mat)
# 这里为了能够更好的提取人脸，手动添加两个特征点
if landmark is None:
return None
add_landmark(landmark)
'''
这部分查看一下人脸的特征变形，这里需要计算人脸剖分的方案，需要覆盖全图
'''
point_dict = from_lanmark_to_points(landmark)
triangles = trianglulation(point_dict)
mask = draw_mask(triangles, point_dict, mat.shape)
img = np.array(mask * mat, dtype=np.uint8)
#img[0] = cv2.equalizeHist(img[0])
#img[1] = cv2.equalizeHist(img[1])
#img[2] = cv2.equalizeHist(img[2])
result = img[landmark["forehead_left"]['y']:landmark["contour_chin"]['y'],landmark["contour_left1"]['x']:landmark["contour_right1"]['x']]
return cv2.resize(result, shape)
def test_face_deform():
"""
对已知剖分结果的三角面进行变形
"""
src = cv2.imread("../../dataset/S076_006_00000001.png")
src = cv2.resize(src, (src.shape[1], src.shape[0]))
dst = cv2.imread("../../dataset/S076_006_00000009.png")
dst = cv2.resize(dst, (src.shape[1], src.shape[0]))
print src.shape
print dst.shape
rows = src.shape[0]
cols = src.shape[1]
seat = np.zeros((rows, cols, 2))
for col in range(cols):
for row in range(rows):
seat[row][col][0] = row
seat[row][col][1] = col
src_landmark = fapi.get_feature_points_fromimage(src)
dst_landmark = fapi.get_feature_points_fromimage(dst)
if dst_landmark is None or src_landmark is None:
print "人脸识别出错"
else:
print "特征点识别成功"
add_landmark(src_landmark)
add_landmark(dst_landmark)
shape = src.shape
add_border_point(src_landmark, (shape[1], shape[0]))
add_border_point(dst_landmark, (shape[1], shape[0]))
'''
这部分查看一下人脸的特征变形，这里需要计算人脸剖分的方案，需要覆盖全图
'''
src_points = from_lanmark_to_points(src_landmark)
dst_points = from_lanmark_to_points(dst_landmark)
triangles = trianglulation(src_points)
affine_mats = compute_affine_mat(triangles, src_points, dst_points)
warped_images = []
warped_seat = []
for M in affine_mats:
I = cv2.warpAffine(src, M, (shape[1], shape[0]))
I_seat = cv2.warpAffine(seat, M, (shape[1], shape[0]))
warped_images.append(I)
warped_seat.append(I_seat)
masks = draw_masks(triangles, dst_points, shape)
mask_seat = draw_masks(triangles, dst_points, seat.shape, color=(1, 1))
result, one = combine_image(warped_images, masks, shape)
result_seat, one_seat = combine_image(warped_seat, mask_seat, seat.shape)
result_img = result + one * src
# draw_triangle(result_img,triangles, src_points)
result = result_seat + one_seat * seat
result[:,:,0] = cv2.normalize(cv2.absdiff(result[:,:,0],seat[:,:,0]),None, 0, 1, cv2.NORM_MINMAX)
result[:,:,1] = cv2.normalize(cv2.absdiff(result[:,:,1],seat[:,:,1]),None, 0, 1, cv2.NORM_MINMAX)
cv2.imshow("dst1", result[:,:,0])
cv2.imshow("dst2", result[:,:,1])
cv2.imshow("dst", result_img/255)
print result
cv2.imshow("src", src)
cv2.waitKey(0)
def camse2_rect_aligened():
"""
将CAMSE2的数据中的人脸，和特征点全部提取出来
:return:
"""
"""`
人脸对齐：
1。读取光流，文件中的内容
2。读取两帧图像
3。读取第一帧图像的特征点
4。计算由光流变到第二帧后的位置。
5。计算仿射变换
6。对第二张图像进行仿射变换来对齐
"""
counter_index = range(
0, 83
) # [ 21, 20, 19, 56, 15, 14, 41, 65, 32, 76, 36, 35,44,68 , 60, 30,46]
path = "D:\\CAME2\\CASME2-coding-20140508.csv"
rect_root = "D:\\CAME2\\Face_Rect\\"
flow_root = "D:\\CAME2\\Flow\\"
aligned_root = "D:\\CAME2\\Face_Rect_Aligned\\"
file = open(path)
lines = file.readlines()
for index, line in enumerate(lines[1:]):
meta = line.split(",")
subject = "sub%02d" % int(meta[0])
if os.path.exists(aligned_root + subject):
pass
else:
os.mkdir(aligned_root + subject)
sqe = meta[1]
if os.path.exists(aligned_root + subject + "\\" + sqe):
pass
else:
os.mkdir(aligned_root + subject + "\\" + sqe)
begin = int(meta[3])
end = int(meta[5])
image1_file_name = rect_root + subject + "\\" + sqe + "\\" + "img" + str(
begin) + ".jpg"
image1 = cv2.imread(image1_file_name)
aligned_name = aligned_root + subject + "\\" + sqe + "\\" + "img" + str(
begin) + ".jpg"
cv2.imwrite(aligned_name, image1[10:-10, 10:-10, :])
vd = fd.get_feature_points_fromimage(image1)
points = np.array(fd.from_lanmark_to_points(vd).values())
feature_pos = points[counter_index]
for i in range(begin + 1, end + 1, 1):
flow_name = flow_root + subject + "\\" + sqe + "\\" + "reg_flow" + str(
i - 1) + ".xml"
image2_file_name = rect_root + subject + "\\" + sqe + "\\" + "img" + str(
i) + ".jpg"
image2 = cv2.imread(image2_file_name)
aligned_name = aligned_root + subject + "\\" + sqe + "\\" + "img" + str(
i) + ".jpg"
if os.path.exists(aligned_name):
continue
U, V = flow_reader(flow_name)
feature_back = []
for pos in feature_pos:
pos[0] = max(pos[0], 0)
pos[1] = max(pos[1], 0)
pos[0] = min(pos[0], image2.shape[1] - 1)
pos[1] = min(pos[1], image2.shape[0] - 1)
new_pos = (pos[0] + U[pos[1]][pos[0]],
pos[1] + V[pos[1]][pos[0]])
feature_back.append(new_pos)
fl = of.OpticalFlow()
mat = fl.calc_affine_mat(np.array(feature_back, dtype=np.float32),
np.array(feature_pos, dtype=np.float32))
B = cv2.warpAffine(image2, mat.T,
(image2.shape[1], image2.shape[0]))
cv2.imwrite(aligned_name, B[10:-10, 10:-10, :])