def test_tf_rgb_lab(img, tf_v1=True):
# raw tensor
raw_input = tf.image.convert_image_dtype(img, dtype=tf.float32)
raw_input.set_shape([None, None, 3])
# convert to lab-space image {L, a, b}
lab = Conv_img.rgb_to_lab(raw_input)
L_chan, a_chan, b_chan = Conv_img.preprocess_lab(lab)
lab = Conv_img.deprocess_lab(L_chan, a_chan, b_chan)
# get back the RGB image (tensor)
true_image = Conv_img.lab_to_rgb(lab)
true_image = tf.image.convert_image_dtype(true_image, dtype=tf.uint8, saturate=True)
# get image array from tensor
if tf_v1: # for tf v1.x
init_op = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init_op)
image = true_image.eval()
else: # for tf v2.0
tf.compat.v1.disable_eager_execution()
init_op = tf.compat.v1.global_variables_initializer()
with tf.compat.v1.Session() as sess:
sess.run(init_op)
image = true_image.numpy()
# save/show image
plt.imshow(image)
plt.imsave('output.jpg', image)
plt.show()
def _augment(
img,
label_img,
resize=None, # Resize the image to some size e.g. [256, 256]
scale=1, # Scale image e.g. 1 / 255.
hue_delta=0, # Adjust the hue of an RGB image by random factor
horizontal_flip=False, # Random left right flip,
vertical_flip=False,
rot=False,
brightness=None,
contrast=None,
saturation=None,
gamma=None,
width_shift_range=0, # Randomly translate the image horizontally
height_shift_range=0,
shape=(256, 256, 3),
cspace='RGB'): # Randomly translate the image vertically
if hue_delta:
img = tf.image.random_hue(img, hue_delta)
img, label_img = flip_img(horizontal_flip, vertical_flip, img, label_img)
img, label_img = shift_img(img, label_img, width_shift_range,
height_shift_range)
img, label_img = rot_img(rot, img, label_img)
if (brightness is not None):
img = tf.image.random_brightness(img, brightness)
if (contrast is not None):
img = tf.image.random_contrast(img, contrast[0], contrast[1])
if (saturation is not None):
img = tf.image.random_saturation(img, saturation[0], saturation[1])
#if (gamma is not None):
# gamma_prob = tf.random_uniform([], 0.0, 1.0)
# img = tf.cond(tf.less(gamma_prob, 0.5),
# lambda: (tf.image.adjust_gamma(img, gamma)),
# lambda: (img))
img = tf.clip_by_value(img, 0.0, 1.0)
label_img = tf.to_float(label_img) * scale
img = tf.to_float(img) * scale
if (cspace == 'RGB'):
img = img
elif (cspace == 'HSV'):
img = tf.image.rgb_to_hsv(img)
elif (cspace == 'RGB-HSV'):
img = tf.concat([img, tf.image.rgb_to_hsv(img)], -1)
elif (cspace == 'LAB'):
img_lab = Conv_img.rgb_to_lab(img)
L, a, b = Conv_img.preprocess_lab(img_lab)
img = tf.stack([L, a, b], axis=2)
elif (cspace == 'RGB-LAB'):
img_lab = Conv_img.rgb_to_lab(img)
L, a, b = Conv_img.preprocess_lab(img_lab)
img_lab = tf.stack([L, a, b], axis=2)
img = tf.concat([img, img_lab], -1)
elif (cspace == 'RGB-HSV-LAB'):
img_rgb_hsv = tf.concat([img, tf.image.rgb_to_hsv(img)], -1)
img_lab = Conv_img.rgb_to_lab(img)
L, a, b = Conv_img.preprocess_lab(img_lab)
img_lab = tf.stack([L, a, b], axis=2)
img = tf.concat([img_rgb_hsv, img_lab], -1)
elif (cspace == 'RGB-HSV-L'):
img_rgb_hsv = tf.concat([img, tf.image.rgb_to_hsv(img)], -1)
img_lab = Conv_img.rgb_to_lab(img)
L, a, b = Conv_img.preprocess_lab(img_lab)
img = tf.concat([img_rgb_hsv, tf.stack([L], axis=2)], -1)
elif (cspace == 'RGB-SV-LAB'):
img_hsv = tf.image.rgb_to_hsv(img)
H, S, V = tf.unstack(img_hsv, axis=2)
img_sv = tf.stack([S, V], axis=2)
img_rgb_sv = tf.concat([img, img_sv], -1)
img_lab = Conv_img.rgb_to_lab(img)
L, a, b = Conv_img.preprocess_lab(img_lab)
img_lab = tf.stack([L, a, b], axis=2)
img = tf.concat([img_rgb_sv, img_lab], -1)
return img, label_img
def _augment(img,
label_img,
cspace='RGB',
augment=True): # Randomly translate the image vertically
seed = np.random.randint(2**32 - 1, dtype=np.int64)
# print('Seed = {}'.format(seed))
if (augment):
img, label_img = jd.transform_img(img,
label_img,
flip_left_right=True,
flip_up_down=True,
crop=0.75,
rot90=True,
brightness=None,
blur=False,
contrast=None,
hue=None,
gamma=None,
saturation=None,
noise=None,
size=(256, 256, 3))
#img = tf.image.random_hue(img, 0.1)
#img, label_img = flip_img(True, img, label_img)
#img, label_img = shift_img(img, label_img, 0.1,
# 0.1)
if (cspace == 'RGB'):
img = img
elif (cspace == 'HSV'):
img = tf.image.rgb_to_hsv(img)
elif (cspace == 'RGB-HSV'):
img = tf.concat([img, tf.image.rgb_to_hsv(img)], -1)
elif (cspace == 'LAB'):
img_lab = Conv_img.rgb_to_lab(img)
L, a, b = Conv_img.preprocess_lab(img_lab)
img = tf.stack([L, a, b], axis=2)
elif (cspace == 'RGB-LAB'):
img_lab = Conv_img.rgb_to_lab(img)
L, a, b = Conv_img.preprocess_lab(img_lab)
img_lab = tf.stack([L, a, b], axis=2)
img = tf.concat([img, img_lab], -1)
elif (cspace == 'RGB-HSV-LAB'):
img_rgb_hsv = tf.concat([img, tf.image.rgb_to_hsv(img)], -1)
img_lab = Conv_img.rgb_to_lab(img)
L, a, b = Conv_img.preprocess_lab(img_lab)
img_lab = tf.stack([L, a, b], axis=2)
img = tf.concat([img_rgb_hsv, img_lab], -1)
elif (cspace == 'RGB-HSV-L'):
img_rgb_hsv = tf.concat([img, tf.image.rgb_to_hsv(img)], -1)
img_lab = Conv_img.rgb_to_lab(img)
L, a, b = Conv_img.preprocess_lab(img_lab)
img = tf.concat([img_rgb_hsv, tf.stack([L], axis=2)], -1)
elif (cspace == 'RGB-SV-LAB'):
img_hsv = tf.image.rgb_to_hsv(img)
H, S, V = tf.unstack(img_hsv, axis=2)
img_sv = tf.stack([S, V], axis=2)
img_rgb_sv = tf.concat([img, img_sv], -1)
img_lab = Conv_img.rgb_to_lab(img)
L, a, b = Conv_img.preprocess_lab(img_lab)
img_lab = tf.stack([L, a, b], axis=2)
img = tf.concat([img_rgb_sv, img_lab], -1)
return img, label_img
parser.add_argument('--im',
required=False,
dest='im',
type=str,
default='data/umn.jpg',
help='image path')
args = parser.parse_args()
img = misc.imread(args.im)
# raw tensor
raw_input = tf.image.convert_image_dtype(img, dtype=tf.float32)
raw_input.set_shape([None, None, 3])
# convert to lab-space image {L, a, b}
lab = Conv_img.rgb_to_lab(raw_input)
L_chan, a_chan, b_chan = Conv_img.preprocess_lab(lab)
lab = Conv_img.deprocess_lab(L_chan, a_chan, b_chan)
# get back the RGB image
true_image = Conv_img.lab_to_rgb(lab)
true_image = tf.image.convert_image_dtype(true_image,
dtype=tf.uint8,
saturate=True)
init_op = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init_op)
# here is the image Tensor :)
image = true_image.eval()
conv_img = Image.fromarray(image, 'RGB')
#conv_img.save('converted_test.jpg')