def __init__(self, background, target):
'''Creates the filter, given background/target means and covariance.
Arguments:
`background` (`GaussianStats`):
The Gaussian statistics for the background (e.g., the result
of calling :func:`calc_stats`).
`target` (ndarray):
Length-K target mean
'''
from math import sqrt
from spectral.algorithms.transforms import LinearTransform
self.background = background
self.u_b = background.mean
self.u_t = target
self._whitening_transform = None
d_tb = (target - self.u_b)
self.d_tb = d_tb
C_1 = background.inv_cov
self.C_1 = C_1
# Normalization coefficient (inverse of squared Mahalanobis distance
# between u_t and u_b)
self.coef = 1.0 / d_tb.dot(C_1).dot(d_tb)
LinearTransform.__init__(
self, (self.coef * d_tb).dot(C_1), pre=-self.u_b)
def __init__(self, background, target):
'''Creates the filter, given background/target means and covariance.
Arguments:
`background` (`GaussianStats`):
The Gaussian statistics for the background (e.g., the result
of calling :func:`calc_stats`).
`target` (ndarray):
Length-K target mean
'''
from math import sqrt
from spectral.algorithms.transforms import LinearTransform
self.background = background
self.u_b = background.mean
self.u_t = target
self._whitening_transform = None
d_tb = (target - self.u_b)
self.d_tb = d_tb
C_1 = background.inv_cov
self.C_1 = C_1
# Normalization coefficient (inverse of squared Mahalanobis distance
# between u_t and u_b)
self.coef = 1.0 / d_tb.dot(C_1).dot(d_tb)
LinearTransform.__init__(self, (self.coef * d_tb).dot(C_1),
pre=-self.u_b)
def __init__(self, transform, img):
from spectral.algorithms.transforms import LinearTransform
if not isinstance(img, Image):
raise Exception(
'Invalid image argument to to TransformedImage constructor.')
if isinstance(transform, numpy.ndarray):
transform = LinearTransform(transform)
self.transform = transform
if self.transform.dim_in not in (None, img.shape[-1]):
raise Exception('Number of bands in image (%d) do not match the '
' input dimension of the transform (%d).' %
(img.shape[-1], transform.dim_in))
params = img.params()
self.set_params(params, params.metadata)
# If img is also a TransformedImage, then just modify the transform
if isinstance(img, TransformedImage):
self.transform = self.transform.chain(img.transform)
self.image = img.image
else:
self.image = img
if self.transform.dim_out is not None:
self.shape = self.image.shape[:2] + (self.transform.dim_out, )
self.nbands = self.transform.dim_out
else:
self.shape = self.image.shape
self.nbands = self.image.nbands
def test_pre_matrix_multiply_post(self):
from spectral.algorithms.transforms import LinearTransform
(i, j, k) = self.datum
transform = LinearTransform(self.matrix, pre=self.pre, post=self.post)
result = transform(self.image[i, j])[k]
assert_almost_equal(result,
self.scalar * (self.pre + self.value) + self.post)
def whiten(self, X):
'''Transforms data to the whitened space of the background.
Arguments:
`X` (ndarray):
Size (M,N,K) or (M*N,K) array of length K vectors to transform.
Returns an array of same size as `X` but linearly transformed to the
whitened space of the filter.
'''
import math
from spectral.algorithms.transforms import LinearTransform
from spectral.algorithms.spymath import matrix_sqrt
if self._whitening_transform is None:
A = math.sqrt(self.coef) * self.background.sqrt_inv_cov
self._whitening_transform = LinearTransform(A, pre=-self.u_b)
return self._whitening_transform(X)
def test_pre_matrix_multiply(self):
(i, j, k) = self.datum
transform = LinearTransform(self.matrix, pre=self.pre)
result = transform(self.image[i, j])[k]
assert_almost_equal(result, self.scalar * (self.pre + self.value))
def test_pre_scalar_multiply_post(self):
(i, j, k) = self.datum
transform = LinearTransform(self.scalar, pre=self.pre, post=self.post)
result = transform(self.image[i, j])[k]
assert_almost_equal(result,
self.scalar * (self.pre + self.value) + self.post)
def test_scalar_multiply(self):
(i, j, k) = self.datum
transform = LinearTransform(self.scalar)
result = transform(self.image[i, j])[k]
assert_almost_equal(result, self.scalar * self.value)
def test_matrix_multiply_post(self):
(i, j, k) = self.datum
transform = LinearTransform(self.matrix, post=self.post)
result = transform(self.image[i, j])[k]
assert_almost_equal(result, self.scalar * self.value + self.post)
def test_scalar_multiply(self):
from spectral.algorithms.transforms import LinearTransform
(i, j, k) = self.datum
transform = LinearTransform(self.scalar)
result = transform(self.image[i, j])[k]
assert_almost_equal(result, self.scalar * self.value)
