def test_number_and_shape_of_scales_match_channels_last(self, num_levels):
nlp = transform.NLP(num_levels=num_levels, data_format="channels_last")
image = tf.zeros((1, 16, 16, 2))
subbands = nlp(image)
self.assertLen(subbands, num_levels)
expected_shapes = [(1, 16, 16, 2), (1, 8, 8, 2)]
for subband, shape in zip(subbands, expected_shapes):
self.assertEqual(subband.shape, shape)
def nlpd(image_a,
image_b,
num_levels=6,
cdm_min=5,
cdm_max=180,
data_format="channels_last"):
"""Normalized Laplacian pyramid distance.
This implements the NLPD as defined in the paper:
> "Perceptually optimized image rendering"</br>
> V. Laparra, A. Berardino, J. Ballé and E. P. Simoncelli</br>
> https://doi.org/10.1364/JOSAA.34.001511
The inputs are assumed to be sRGB images, and the display is assumed to have a
dynamic range of `cdm_min` to `cdm_max` cd/m^2.
Args:
image_a: `Tensor` containing image A.
image_b: `Tensor` containing image B.
num_levels: Integer. The number of pyramid levels, including the lowpass
residual.
cdm_min: Float. Minimum assumed cd/m^2 of display.
cdm_max: Float. Maximum assumed cd/m^2 of display.
data_format: String. Either `'channels_first'` or `'channels_last'`.
Returns:
A `Tensor` giving the NLPD values for each of the non-spatial dimensions
(e.g. shaped NC for NHWC inputs).
"""
if not 0 <= cdm_min < cdm_max:
raise ValueError("Must have `0 <= cdm_min < cdm_max`.")
if image_a.dtype.is_integer:
image_a = tf.cast(image_a, tf.float32)
if image_b.dtype.is_integer:
image_b = tf.cast(image_b, tf.float32)
def convert_to_cdm2(image):
return ((image / 255)**2.4) * (cdm_max - cdm_min) + cdm_min
nlp = transform.NLP(num_levels=num_levels, data_format=data_format)
subbands_a = nlp(convert_to_cdm2(image_a))
subbands_b = nlp(convert_to_cdm2(image_b))
if data_format == "channels_first":
spatial_axes = (-2, -1)
else:
spatial_axes = (-3, -2)
return lp_norm(subbands_a, subbands_b, spatial_axes=spatial_axes)
def nlpd_fast(image_a, image_b, num_levels=6, data_format="channels_last"):
"""Normalized Laplacian pyramid distance.
This implements a quick-and-dirty approximation to the NLPD, which is defined
in the paper:
> "Perceptually optimized image rendering"</br>
> V. Laparra, A. Berardino, J. Ballé and E. P. Simoncelli</br>
> https://doi.org/10.1364/JOSAA.34.001511
The inputs are assumed to be sRGB images. This approximation omits the
colorspace conversion.
Args:
image_a: `Tensor` containing image A.
image_b: `Tensor` containing image B.
num_levels: Integer. The number of pyramid levels, including the lowpass
residual.
data_format: String. Either `'channels_first'` or `'channels_last'`.
Returns:
A `Tensor` giving the NLPD values for each of the non-spatial dimensions
(e.g. shaped NC for NHWC inputs).
"""
if image_a.dtype.is_integer:
image_a = tf.cast(image_a, tf.float32)
if image_b.dtype.is_integer:
image_b = tf.cast(image_b, tf.float32)
nlp = transform.NLP(num_levels=num_levels,
gamma=None,
data_format=data_format)
subbands_a = nlp(image_a)
subbands_b = nlp(image_b)
if data_format == "channels_first":
spatial_axes = (-2, -1)
else:
spatial_axes = (-3, -2)
return lp_norm(subbands_a, subbands_b, spatial_axes=spatial_axes)
def test_invalid_shapes_fail_channels_last(self):
nlp = transform.NLP(data_format="channels_last")
with self.assertRaises(ValueError):
nlp.build((16, 16))
def test_invalid_data_format_fails(self):
with self.assertRaises(ValueError):
transform.NLP(data_format=3)
def test_invalid_gamma_fails(self):
with self.assertRaises(ValueError):
transform.NLP(gamma=-1)
def test_invalid_num_levels_fails(self, num_levels):
with self.assertRaises(ValueError):
transform.NLP(num_levels=num_levels)
