def update_particle_field(self, poses=None, add=True):
if poses is None:
poses = self.pos
wid = self.viewrad
for p in poses:
#1. get tile
l = np.clip(p - 2 * wid, 0, self.image.shape)
r = np.clip(p + 2 * wid, 0, self.image.shape)
t = Tile(l, r, mins=0, maxs=self.image.shape)
#2. update:
c = t.coords(form='broadcast')
if add:
self.particle_field[t.slicer] += self._particle_func(c, p, wid)
else:
self.particle_field[t.slicer] -= self._particle_func(c, p, wid)
def get_update_tile(self, params, values):
""" Get the amount of support size required for a particular update."""
doglobal, particles = self._update_type(params)
if doglobal:
return self.shape.copy()
# 1) store the current parameters of interest
values0 = self.get_values(params)
# 2) calculate the current tileset
tiles0 = [self._tile(n) for n in particles]
# 3) update to newer parameters and calculate tileset
self.set_values(params, values)
tiles1 = [self._tile(n) for n in particles]
# 4) revert parameters & return union of all tiles
self.set_values(params, values0)
return Tile.boundingtile(tiles0 + tiles1)
def _tile(self, n):
""" Get the tile surrounding particle `n` """
zsc = np.array([1.0 / self.zscale, 1, 1])
pos, rad = self.pos[n], self.rad[n]
pos = self._trans(pos)
return Tile(pos - zsc * rad, pos + zsc * rad).pad(self.support_pad)
def _tile(self, n):
"""Get the update tile surrounding particle `n` """
pos = self._trans(self.pos[n])
return Tile(pos, pos).pad(self.support_pad)
def get_padding_size(self, tile, z=None):
return Tile(self.support)
def get_padding_size(self, tile):
self.px = np.sqrt(-2 * np.log(self.error) * self.values[2]**2)
self.py = np.sqrt(-2 * np.log(self.error) * self.values[1]**2)
self.pz = np.sqrt(-2 * np.log(self.error) * self.values[0]**2)
return Tile(np.ceil([self.pz, self.py, self.px]))
def get_padding_size(self, tile):
return Tile(np.ones(3))
def _rvecs(self, shape, centered=True):
tile = Tile(shape)
return tile.kvectors(norm=1.0 / tile.shape, shift=centered)
def identify_misfeatured_regions(st, filter_size=5, sigma_cutoff=8.):
"""
Identifies regions of missing/misfeatured particles based on the
residuals' local deviation from uniform Gaussian noise.
Parameters
----------
st : :class:`peri.states.State`
The state in which to identify mis-featured regions.
filter_size : Int, best if odd.
The size of the filter for calculating the local standard deviation;
should approximately be the size of a poorly featured region in
each dimension. Default is 5.
sigma_cutoff : Float or `otsu`, optional
The max allowed deviation of the residuals from what is expected,
in units of the residuals' standard deviation. Lower means more
sensitive, higher = less sensitive. Default is 8.0, i.e. one pixel
out of every 7*10^11 is mis-identified randomly. In practice the
noise is not Gaussian so there are still some regions mis-identified
as improperly featured. Set to ```otsu``` to calculate this number
based on an automatic Otsu threshold.
Returns
-------
tiles : List of :class:`peri.util.Tile`
Each tile is the smallest bounding tile that contains an improperly
featured region. The list is sorted by the tile's volume.
Notes
-----
Algorithm is
1. Create a field of the local standard deviation, as measured over
a hypercube of size filter_size.
2. Find the maximum reasonable value of the field. [The field should
be a random variable with mean of r.std() and standard deviation
of ~r.std() / sqrt(N), where r is the residuals and N is the
number of pixels in the hypercube.]
3. Label & Identify the misfeatured regions as portions where
the local error is too large.
4. Parse the misfeatured regions into tiles.
5. Return the sorted tiles.
The Otsu option to calculate the sigma cutoff works well for images
that actually contain missing particles, returning a number similar
to one calculated with a sigma cutoff. However, if the image is
well-featured with Gaussian residuals, then the Otsu threshold
splits the Gaussian down the middle instead of at the tails, which
is very bad. So use with caution.
"""
# 1. Field of local std
r = st.residuals
weights = np.ones([filter_size] * len(r.shape), dtype='float')
weights /= weights.sum()
f = np.sqrt(nd.filters.convolve(r * r, weights, mode='reflect'))
# 2. Maximal reasonable value of the field.
if sigma_cutoff == 'otsu':
max_ok = initializers.otsu_threshold(f)
else:
# max_ok = f.mean() * (1 + sigma_cutoff / np.sqrt(weights.size))
max_ok = f.mean() + sigma_cutoff * f.std()
# 3. Label & Identify
bad = f > max_ok
labels, n = nd.measurements.label(bad)
inds = []
for i in range(1, n + 1):
inds.append(np.nonzero(labels == i))
# 4. Parse into tiles
tiles = [
Tile(np.min(ind, axis=1),
np.max(ind, axis=1) + 1) for ind in inds
]
# 5. Sort and return
volumes = [t.volume for t in tiles]
return [tiles[i] for i in np.argsort(volumes)[::-1]]
# im = io.imread('/Volumes/PhD/DavidData/Emily/2014_5_23-T3Y_63xoil_1.lsm', as_gray=True)
# im_array = np.array(im)
# im.shape
# im_array_small = im_array[5:,312:, 0:200]
# im_array_small.shape
# f=tp.locate(im_array,diameter=(3,11,11))
# f[f['z']==f['z']]
# f_small = tp.locate(im_array_small,diameter=(3,11,11))
# f_small=f_small[f_small['z']==f_small['z']]
# np.save('part_loc_T3Y_1.npy', np.array(f[f.columns[0:3,]]))
imFile = '/Volumes/PhD/DavidData/Emily/2014_5_23-T3Y_63xoil_1.lsm'
raw_im = RawImage(imFile)
im_arr = raw_im.get_image()
im_arr.shape()
f = tp.locate(im_arr, diameter=(9, 9, 9))
tile = Tile(200)
small_im = RawImage(imFile, tile=tile)
small_im_arr = small_im.get_image()
f_small = tp.locate(small_im_arr, diameter=(9, 9, 9))
particle_positions = np.array(f_small[f_small.columns[0:3, ]])
np.save('part_loc_T3Y_1.npy', particle_positions)
# plt.figure()
# tp.annotate3d(f[f['z']==f['z']],im_array)
plt.imshow(im_array[2, :, :], cmap="gray")
scatter = plt.scatter(x=f["x"], y=f["y"], c=f["z"], marker="x")
plt.legend(*scatter.legend_elements(), title="z")
plt.show()
