def segment_image_from_labeled_modes(
image, labels, cell_edges, spatial_features=True
):
# for each pixel in the image, look up the (x,y,c1,c2,c3) feature in
# the labels function -- this is just a matter of downsampling
# (or dividing the indices)
# features correspond to pixels in row-major order,
# so we can just zip the whole
# course-grid feature indices into a flat list
features = image_to_features(image)
if not spatial_features:
features = features[:,2:]
cell_idc = nearest_cell_idx(features, cell_edges)
flat_idc = np.lib.index_tricks.ravel_multi_index(
cell_idc.T, labels.shape
)
seg_img = labels.flat[flat_idc]
seg_img.shape = image.shape[:2]
return seg_img
def resolve_label_boundaries(
labels, grid, cell_edges, max_iter = 20
):
"""
For each boundary point in labels, perform a gradient ascent using
the mean shift sequence. Terminate the walk once the sequence enters
a region already marked as belonging to a mode.
Parameters
----------
labels: ndarray
label map of the histogram grid points
p: ndarray
density over grid
cell_edges: sequence
bin edges of the histogram (needed for grid index conversion)
features: ndarray
the feature vectors of the original image
sigma: float
kernel bandwidth
"""
# the boundary points should be ordered by decreasing density
## boundary_pts = np.where(labels.ravel()==boundary)[0]
boundary_pts = np.where(labels.ravel()<0)[0]
p = grid[...,-1]
order = np.argsort(p.flat[boundary_pts])[::-1]
## order = np.argsort(p.flat[boundary_pts])[::-1]
boundary_pts = np.unravel_index(boundary_pts[order], labels.shape)
## grid = np.concatenate( (mu, p[...,None]), axis=-1).copy()
## boundary_pts = np.unravel_index(boundary_pts, labels.shape)
gdims = grid.shape[:-1]
prev_ms = np.zeros((len(gdims),), 'd')
unit_edges = [np.arange(d+1, dtype='d') for d in gdims]
for x in itertools.izip(*boundary_pts):
assigned = False
fvec = np.array(x, 'd') + 0.5
n_iter = 0
prev_ms[:] = 0
while not assigned and n_iter < max_iter:
mean = multilinear_interpolation(fvec, grid)
density = mean[-1]
mean = mean[:-1]
if density > 1e-4:
mean /= density
else:
assigned = True
ms = mean - fvec
cs = np.dot(ms, prev_ms)
if cs < 0:
ms -= prev_ms * cs
norm_sq = np.dot(ms, ms)
if norm_sq > 1e-8:
if norm_sq < (0.2**2):
ms *= 0.2/np.sqrt(norm_sq)
fvec += ms
prev_ms = ms/np.sqrt(norm_sq)
else:
assigned = True
idx = nearest_cell_idx(fvec[None,:], unit_edges)[0]
label = labels[tuple(idx)]
if label > 0:
labels[x] = label
assigned = True
n_iter += 1
def resolve_segmentation_boundaries(
seg_image, features, grid, labels, cell_edges, max_iter = 20
):
# NOTE: grid is now a ND manifold whose
# range concatenates grid means and density
bpts = np.where(seg_image<=0)
ishape = seg_image.shape
k = 1
n_pt = len(bpts[0])
next_pct = 10
idx = np.zeros((features.shape[-1],), 'l')
## max_idx = np.array(grid.shape[:len(idx)], 'l')
## max_idx -= 1
gdims = grid.shape[:-1]
prev_ms = np.zeros((len(gdims),), 'd')
unit_edges = [np.arange(d+1, dtype='d') for d in gdims]
for y, x in itertools.izip(*bpts):
pct = int(100*(k/float(n_pt)))
k += 1
if pct==next_pct:
next_pct += 10
print pct, '%'
fvec = normalized_feature(features[y*ishape[1] + x], cell_edges)
assigned = False
n_iter = 0
prev_ms[:] = 0
while not assigned and n_iter < max_iter:
mean = multilinear_interpolation(fvec, grid)
density = mean[-1]
## print n_iter, density
mean = mean[:-1]
if density > 1e-4:
mean /= density
else:
assigned = True
ms = mean - fvec
## fvec = mean
cs = np.dot(ms, prev_ms)
if cs < 0:
ms -= prev_ms * cs
norm_sq = np.dot(ms, ms)
if norm_sq > 1e-8:
if norm_sq < (0.2**2):
ms *= 0.2/np.sqrt(norm_sq)
fvec += ms
prev_ms = ms/np.sqrt(norm_sq)
else:
assigned = True
idx = nearest_cell_idx(fvec[None,:], unit_edges)[0]
label = labels[tuple(idx)]
if label > 0:
seg_image[y,x] = label
assigned = True
n_iter += 1
## print 'label', label, 'after', n_iter, 'iters'
return
