def load_data_val_gray(data,
pathX,
img_rows=224,
img_cols=224,
channels=4,
file_name_suffix=None,
standardize=False,
normalize=True,
contrast=False,
equalize=False,
stretch_img=False,
data_augmentation=False,
angle_norm_method=1,
max_angle=90.0):
X = np.zeros((len(data), channels, img_rows, img_cols))
Y = np.zeros((len(data), 3))
last_id = 0
last_frame = 0
ridx = 0
for i, d in enumerate(data):
folder = "%02d/" % d['id'] + "/"
if file_name_suffix is None:
file = "frame_" + d['frame'] + "_face_gray.png"
else:
file = "frame_" + d['frame'] + file_name_suffix + ".png"
if channels > 1:
img = cv2.imread(pathX + folder + file, 1)
else:
img = cv2.imread(pathX + folder + file, 0)
if img is None:
continue
if data_augmentation:
if last_id == int(d['id']) and last_frame == int(d['frame']):
ridx += 1
else:
ridx = 0
rdim = 14
if ridx == 1:
img = img[:img.shape[0] - rdim, :img.shape[1] - rdim]
elif ridx == 2:
img = img[:img.shape[0] - rdim, rdim:img.shape[1]]
elif ridx == 3:
img = img[rdim:img.shape[0], :img.shape[1] - rdim]
elif ridx == 4:
img = img[rdim:img.shape[0], rdim:img.shape[1]]
elif ridx == 5:
img = img[int(rdim / 2):img.shape[0] - int(rdim / 2),
int(rdim / 2):img.shape[1] - int(rdim / 2)]
if stretch_img:
img = stretch(img, 190)
if equalize:
cv2.equalizeHist(img.astype(np.uint8), img)
if contrast:
img = img.astype(np.float32)
if channels > 1:
for i_ch in range(channels):
img[..., i_ch] = exposure.rescale_intensity(
img[..., i_ch].astype('float'), out_range=(0, 1))
else:
img = exposure.rescale_intensity(img.astype('float'),
out_range=(0, 1))
if standardize:
img = img.astype(np.float32)
if channels > 1:
for i_ch in range(channels):
img[...,
i_ch] = preprocessing.scale(img[...,
i_ch].astype('float'))
else:
img = preprocessing.scale(img.astype('float'))
img = img.astype(np.float32)
if img.shape[1] != img_cols or img.shape[0] != img_rows:
img = cv2.resize(img, (img_cols, img_rows))
if channels == 1:
img = np.expand_dims(img, axis=0)
else:
img = np.reshape(img, (img.shape[2], img.shape[0], img.shape[1]))
X[i] = img
if angle_norm_method == 1:
Y[i, 0] = (d['angle1'] + max_angle) / (2.0 * max_angle)
Y[i, 1] = (d['angle2'] + max_angle) / (2.0 * max_angle)
Y[i, 2] = (d['angle3'] + max_angle) / (2.0 * max_angle)
elif angle_norm_method == 0:
Y[i, 0] = (d['angle1'])
Y[i, 1] = (d['angle2'])
Y[i, 2] = (d['angle3'])
elif angle_norm_method == -1:
Y[i, 0] = d['angle1'] / float(max_angle)
Y[i, 1] = d['angle2'] / float(max_angle)
Y[i, 2] = d['angle3'] / float(max_angle)
else:
raise ValueError('Flag: "angle_norm_method" is out of range.')
last_id = int(d['id'])
last_frame = int(d['frame'])
return X, Y
def load_data_gray(data,
pathX,
img_rows=224,
img_cols=224,
channels=3,
file_name_suffix=None,
standardize=False,
contrast=False,
equalize=False,
stretch_img=False,
data_augmentation=True,
angle_norm_method=1,
max_angle=90.0):
X = np.zeros((len(data), channels, img_rows, img_cols))
Y = np.zeros((len(data), 3))
for i, d in enumerate(data):
folder = "%02d/" % d['id'] + "/"
if file_name_suffix is None:
file = "frame_" + d['frame'] + "_face_gray.png"
else:
file = "frame_" + d['frame'] + file_name_suffix + ".png"
if channels > 1:
img = cv2.imread(pathX + folder + file, 1)
else:
img = cv2.imread(pathX + folder + file, 0)
if img is None:
continue
if data_augmentation:
data_augmentation_idx = d['data_augmentation']
rdim = random.randrange(1, 15)
# Corner Top-left
if data_augmentation_idx == 1:
img = img[:img.shape[0] - rdim, :img.shape[1] - rdim]
# Corner Top-right
elif data_augmentation_idx == 2:
img = img[:img.shape[0] - rdim, rdim:img.shape[1]]
# Corner Bottom-left
elif data_augmentation_idx == 3:
img = img[rdim:img.shape[0], :img.shape[1] - rdim]
# Corner bottom-right
elif data_augmentation_idx == 4:
img = img[rdim:img.shape[0], rdim:img.shape[1]]
# Center
elif data_augmentation_idx == 5:
img = img[int(rdim / 2):img.shape[0] - int(rdim / 2),
int(rdim / 2):img.shape[1] - int(rdim / 2)]
# Top
elif data_augmentation_idx == 6:
img = img[int(rdim / 2):, :]
# Bottom
elif data_augmentation_idx == 7:
img = img[:img.shape[0] - int(rdim / 2), :]
# Left
elif data_augmentation_idx == 8:
img = img[:, int(rdim / 2):]
# Right
elif data_augmentation_idx == 9:
img = img[:, :img.shape[1] - int(rdim / 2)]
elif data_augmentation_idx == 10:
noise = np.zeros_like(img, 'float') + np.random.normal(
0.0, 5, img.shape)
img = (img + noise).astype(np.uint8)
img[img < 0] = 0
img[img > 255] = 255
if stretch_img:
img = stretch(img, 190)
if equalize:
cv2.equalizeHist(img.astype(np.uint8), img)
if contrast:
img = img.astype(np.float32)
if channels > 1:
for i_ch in range(channels):
img[..., i_ch] = exposure.rescale_intensity(
img[..., i_ch].astype('float'), out_range=(0, 1))
else:
img = exposure.rescale_intensity(img.astype('float'),
out_range=(0, 1))
if standardize:
img = img.astype(np.float32)
if channels > 1:
for i_ch in range(channels):
img[...,
i_ch] = preprocessing.scale(img[...,
i_ch].astype('float'))
else:
img = preprocessing.scale(img.astype('float'))
img = img.astype(np.float32)
if img.shape[1] != img_cols or img.shape[0] != img_rows:
img = cv2.resize(img, (img_cols, img_rows))
if channels == 1:
img = np.expand_dims(img, axis=0)
else:
img = np.reshape(img, (img.shape[2], img.shape[0], img.shape[1]))
X[i] = img
if angle_norm_method == 1:
Y[i, 0] = (d['angle1'] + max_angle) / (2.0 * max_angle)
Y[i, 1] = (d['angle2'] + max_angle) / (2.0 * max_angle)
Y[i, 2] = (d['angle3'] + max_angle) / (2.0 * max_angle)
elif angle_norm_method == 0:
Y[i, 0] = (d['angle1'])
Y[i, 1] = (d['angle2'])
Y[i, 2] = (d['angle3'])
elif angle_norm_method == -1:
Y[i, 0] = d['angle1'] / float(max_angle)
Y[i, 1] = d['angle2'] / float(max_angle)
Y[i, 2] = d['angle3'] / float(max_angle)
else:
raise ValueError('Flag: "angle_norm_method" is out of range.')
return X, Y