def gm2Gradient(gm, sigma):
"""'reverse-engineer' gradient vector image from scalar image
via Hessian from Gaussian derivative filters of given scale"""
# create vectors in direction of negative curvature:
er = vigra.tensorEigenRepresentation(vigra.gaussianHessian(gm, sigma))
return vigra.transformImage(er, gm,
"\l e,i: i*Vector(-sin(e[2]), cos(e[2]))")
def gm2Gradient(gm, sigma):
"""'reverse-engineer' gradient vector image from scalar image
via Hessian from Gaussian derivative filters of given scale"""
# create vectors in direction of negative curvature:
er = vigra.tensorEigenRepresentation(
vigra.gaussianHessian(gm, sigma))
return vigra.transformImage(
er, gm, "\l e,i: i*Vector(-sin(e[2]), cos(e[2]))")
def colorGradient(img, scale, sqrt = True):
"""Calculate Gaussian color gradient. Calculates sum of Gaussian
gradient tensors in all single bands, and returns pair of (gm,
grad) where gm is the square root of the trace of the color
gradient tensor and grad is a 2-band image with the large
eigenvectors of the appropriate sqrt(gm2) lengths.
If `sqrt` is set to False, returns (gm2, grad) instead, i.e. does
not apply sqrt to the first image (slight optimization if you
don't need the gm image)."""
colorTensor = gradientTensor(img, scale)
# gm2 := sum of squared magnitudes of grad. in each channel
gm2 = vigra.tensorTrace(colorTensor)
# rebuild vector in dir. of large eigenvector with length sqrt(gm2):
grad = vigra.transformImage( # FIXME: use tensor2Gradient?
vigra.tensorEigenRepresentation(colorTensor), gm2,
"\l e, mag2: sqrt(mag2)*Vector(cos(-e[2]), sin(-e[2]))", {})
if sqrt:
gm = vigra.transformImage(gm2, "\l x: sqrt(x)")
return gm, grad
return gm2, grad
def colorGradient(img, scale, sqrt=True):
"""Calculate Gaussian color gradient. Calculates sum of Gaussian
gradient tensors in all single bands, and returns pair of (gm,
grad) where gm is the square root of the trace of the color
gradient tensor and grad is a 2-band image with the large
eigenvectors of the appropriate sqrt(gm2) lengths.
If `sqrt` is set to False, returns (gm2, grad) instead, i.e. does
not apply sqrt to the first image (slight optimization if you
don't need the gm image)."""
colorTensor = gradientTensor(img, scale)
# gm2 := sum of squared magnitudes of grad. in each channel
gm2 = vigra.tensorTrace(colorTensor)
# rebuild vector in dir. of large eigenvector with length sqrt(gm2):
grad = vigra.transformImage( # FIXME: use tensor2Gradient?
vigra.tensorEigenRepresentation(colorTensor), gm2,
"\l e, mag2: sqrt(mag2)*Vector(cos(-e[2]), sin(-e[2]))", {})
if sqrt:
gm = vigra.transformImage(gm2, "\l x: sqrt(x)")
return gm, grad
return gm2, grad
def tensor2Gradient(tensor):
"""rebuild vectors in dir. of large eigenvectors with lengths sqrt(trace)"""
return vigra.transformImage(
vigra.tensorEigenRepresentation(tensor),
"\l e: sqrt(e[0]+e[1])*Vector(cos(-e[2]), sin(-e[2]))", {})
def tensor2Gradient(tensor):
"""rebuild vectors in dir. of large eigenvectors with lengths sqrt(trace)"""
return vigra.transformImage(
vigra.tensorEigenRepresentation(tensor),
"\l e: sqrt(e[0]+e[1])*Vector(cos(-e[2]), sin(-e[2]))", {})
