def label_particles_edge(im, sigma=2, closing_size=0, **extra_args):
""" Segment image using Canny edge-finding filter.
parameters
----------
im : image in which to find particles
sigma : size of the Canny filter
closing_size : size of the closing filter
returns
-------
labels : an image array of uniquely labeled segments
"""
from skimage.morphology import square, binary_closing, skeletonize
if skimage_version < StrictVersion('0.11'):
from skimage.filter import canny
else:
from skimage.filters import canny
edges = canny(im, sigma=sigma)
if closing_size > 0:
edges = binary_closing(edges, square(closing_size))
edges = skeletonize(edges)
labels = sklabel(edges)
print "found {} segments".format(labels.max())
# in ma.array mask, False is True, and vice versa
labels = np.ma.array(labels, mask=edges == 0)
return labels
def label_particles_edge(im, sigma=2, closing_size=0, **extra_args):
""" label_particles_edge(image, sigma=3, closing_size=3)
Returns the labels for an image.
Segments using Canny edge-finding filter.
keyword arguments:
image -- The image in which to find particles
sigma -- The size of the Canny filter
closing_size -- The size of the closing filter
"""
from skimage.morphology import square, binary_closing, skeletonize
if skversion < version('0.11'):
from skimage.filter import canny
else:
from skimage.filters import canny
edges = canny(im, sigma=sigma)
if closing_size > 0:
edges = binary_closing(edges, square(closing_size))
edges = skeletonize(edges)
labels = sklabel(edges)
print "found {} segments".format(labels.max())
labels = np.ma.array(
labels,
mask=edges == 0) # in ma.array mask, False is True, and vice versa
return labels
def label_particles_edge(im, sigma=2, closing_size=0, **extra_args):
""" Segment image using Canny edge-finding filter.
parameters
----------
im : image in which to find particles
sigma : size of the Canny filter
closing_size : size of the closing filter
returns
-------
labels : an image array of uniquely labeled segments
"""
from skimage.morphology import square, binary_closing, skeletonize
if skimage_version < StrictVersion('0.11'):
from skimage.filter import canny
else:
from skimage.filters import canny
edges = canny(im, sigma=sigma)
if closing_size > 0:
edges = binary_closing(edges, square(closing_size))
edges = skeletonize(edges)
labels = sklabel(edges)
print "found {} segments".format(labels.max())
# in ma.array mask, False is True, and vice versa
labels = np.ma.array(labels, mask=edges == 0)
return labels
def label_particles_convolve(im, thresh=3, rmv=None, kern=0, **extra_args):
""" label_particles_convolve(im, thresh=2)
Returns the labels for an image
Segments using a threshold after convolution with proper gaussian kernel
Uses center of mass from original image to find centroid of segments
Input:
image the original image
thresh the threshold above which pixels are included
if integer, in units of intensity std dev
if float, in absolute units of intensity
rmv if given, the positions at which to remove large dots
kern kernel size
"""
# Michael removed disks post-convolution
if rmv is not None:
im = remove_disks(im, *rmv)
if kern == 0:
raise ValueError('kernel size `kern` not set')
kernel = np.sign(kern)*gdisk(abs(kern))
convolved = convolve(im, kernel)
convolved -= convolved.min()
convolved /= convolved.max()
if isinstance(thresh, int):
if rmv is not None:
thresh -= 1 # smaller threshold for corners
thresh = convolved.mean() + thresh*convolved.std()
labels = sklabel(convolved > thresh)
#print "found {} segments above thresh".format(labels.max())
return labels, convolved
def label_particles_convolve(im, kern, thresh=3, rmv=None, **extra_args):
""" Segment image using convolution with gaussian kernel and threshold
parameters
----------
im : the original image to be labeled
kern : kernel size
thresh : the threshold above which pixels are included
if thresh >= 1, in units of intensity std dev
if thresh < 1, in absolute units of intensity
rmv : if given, the positions at which to remove large dots
returns
-------
labels : an image array of uniquely labeled segments
convolved : the convolved image before thresholding and segementation
"""
if rmv is not None:
im = remove_disks(im, *rmv)
kernel = np.sign(kern) * gdisk(abs(kern) / 4, abs(kern))
convolved = ndimage.convolve(im, kernel)
convolved -= convolved.min()
convolved /= convolved.max()
if args.plot > 2:
snapshot('kern', kernel)
snapshot('convolved', convolved, cmap='gray')
if thresh >= 1:
if rmv is not None:
thresh -= 1 # smaller threshold for corners
thresh = thresh * convolved.std() + convolved.mean()
threshed = convolved > thresh
labels = sklabel(threshed, connectivity=1)
if args.plot > 2:
snapshot('threshed', threshed)
snapshot('labeled', np.where(labels, labels, np.nan), cmap='prism_r')
return labels, convolved
def label_particles_convolve(im, kern, thresh=3, rmv=None, **extra_args):
""" Segment image using convolution with gaussian kernel and threshold
parameters
----------
im : the original image to be labeled
kern : kernel size
thresh : the threshold above which pixels are included
if thresh >= 1, in units of intensity std dev
if thresh < 1, in absolute units of intensity
rmv : if given, the positions at which to remove large dots
returns
-------
labels : an image array of uniquely labeled segments
convolved : the convolved image before thresholding and segementation
"""
if rmv is not None:
im = remove_disks(im, *rmv)
kernel = np.sign(kern)*gdisk(abs(kern)/4, abs(kern))
convolved = ndimage.convolve(im, kernel)
convolved -= convolved.min()
convolved /= convolved.max()
if args.plot > 2:
snapshot('kern', kernel)
snapshot('convolved', convolved, cmap='gray')
if thresh >= 1:
if rmv is not None:
thresh -= 1 # smaller threshold for corners
thresh = thresh*convolved.std() + convolved.mean()
threshed = convolved > thresh
labels = sklabel(threshed, connectivity=1)
if args.plot > 2:
snapshot('threshed', threshed)
snapshot('labeled', np.where(labels, labels, np.nan), cmap='prism_r')
return labels, convolved
def label_particles_edge(im, sigma=2, closing_size=0, **extra_args):
""" label_particles_edge(image, sigma=3, closing_size=3)
Returns the labels for an image.
Segments using Canny edge-finding filter.
keyword arguments:
image -- The image in which to find particles
sigma -- The size of the Canny filter
closing_size -- The size of the closing filter
"""
from skimage.morphology import square, binary_closing, skeletonize
if skversion < version('0.11'):
from skimage.filter import canny
else:
from skimage.filters import canny
edges = canny(im, sigma=sigma)
if closing_size > 0:
edges = binary_closing(edges, square(closing_size))
edges = skeletonize(edges)
labels = sklabel(edges)
print "found {} segments".format(labels.max())
labels = np.ma.array(labels, mask=edges==0) # in ma.array mask, False is True, and vice versa
return labels
def label_particles_convolve(im, thresh=3, rmv=None, csize=0, **extra_args):
""" label_particles_convolve(im, thresh=2)
Returns the labels for an image
Segments using a threshold after convolution with proper gaussian kernel
Uses center of mass from original image to find centroid of segments
Input:
image the original image
thresh the threshold above which pixels are included
if integer, in units of intensity std dev
if float, in absolute units of intensity
rmv if given, the positions at which to remove large dots
csize kernel size
"""
# Michael removed disks post-convolution
if rmv is not None:
im = remove_disks(im, *rmv)
if csize == 0:
raise ValueError('kernel size `csize` not set')
elif csize < 0:
ckern = -gdisk(-csize)
else:
ckern = gdisk(csize)
convolved = convolve(im, ckern)
convolved -= convolved.min()
convolved /= convolved.max()
if isinstance(thresh, int):
if rmv is not None:
thresh -= 1 # smaller threshold for corners
thresh = convolved.mean() + thresh * convolved.std()
labels = sklabel(convolved > thresh)
#print "found {} segments above thresh".format(labels.max())
return labels, convolved
def evaluate_nms_results_overlap(test_data,
test_labels,
test_predictions,
print_steps=False,
orig_only=False,
mode="argmax",
Pmin=0.5,
padding=0,
sliding_window_size=(48, 48),
from_config=False,
config={},
min_ac=0.4,
min_acc=0.6):
""" Ohodnoti prekryti vyslednych bounding boxu s artefakty """
# inicializace statistik
TP, TN, FP, FN = 0, 0, 0, 0
problematic = list()
bounding_boxes = get_boxes(test_predictions,
test_labels,
padding=padding,
from_config=from_config,
config=config)
#print(bounding_boxes)
for index in range(test_labels.shape[0]):
mask = test_labels[index].astype("uint8") * 255
boxes = bounding_boxes[index]
mask = np.argmax(mask, axis=2) #*127
TP0, TN0, FP0, FN0 = 0, 0, 0, 0
# oriznuti obrazku a masky -> takhle se to dela u augmentovanych
#img, mask = fe.cut_image(orig, mask)
#mask /= 127.0
# olabelovani artefaktu
imlabel = sklabel(mask)
# obarveni mist bez artefaktu na 0
imlabel[(mask == 0) | (mask == 2)] = 0
# vytvoreni prazdneho obrazku
blank = np.zeros(mask.shape)
# ziskani indexu artefaktu
artefact_ids = np.unique(imlabel)[1:]
# seznam boxu, ktere pokryvaji nejaky artefakt
covered_box_ids = list()
# prochazeni vsech artefaktu
for i in artefact_ids:
covered_by_bb = False
for j, (y, h, x, w) in enumerate(boxes):
# obarveni oblasti boxu
blank[y:h, x:w] = 1
# vypocet pixelu artefaktu celeho a v boxu
na = np.sum((imlabel == i).astype(int))
nab = np.sum((imlabel == i) & (blank == 1))
# vypocet zastoupeni bb v artefaktu
artefact_bb_coverage = float(nab) / na
# pokud je artefakt alespon z poloviny pokryt boxem
if artefact_bb_coverage >= 0.5:
#print artefact_bb_coverage
covered_box_ids.append(j)
# vytazeni frmau masky
mask_frame = mask[y:h, x:w]
# pokud jsou pokryty artefaktem -> TP, jinak FP
if covered_by_artefact(mask_frame,
min_ac=min_ac,
min_acc=min_acc):
TP += 1
TP0 += 1
covered_by_bb = True
break
elif artefact_bb_coverage == 1.0 and (np.sum(
(blank == 1).astype(int)) == sliding_window_size[0]**
2):
TP += 1
TP0 += 1
covered_by_bb = True
break
else:
FP += 1
FP0 += 1
# znovu prebarveni pomocneho framu zpatky na 0
blank[y:h, x:w] = 0
# pokud neni pokryt zadnym boxem alespon z poloviny
if not covered_by_bb: # and na >= 300: # navic by mel byt dostatecne velky
FN += 1
FN0 += 1
# prochazeni zatim neprohlendutych boxu
for j in range(len(boxes)):
if not j in covered_box_ids:
FP += 1
FP0 += 1
"""
# vytazeni boxu
y, h, x, w = boxes[j]
mask_frame = mask[y:h, x:w]
# pokud jsou pokryty artefaktem -> TP, jinak FP
if covered_by_artefact(mask_frame):
TP += 1
TP0 += 1
else:
FP += 1
FP0 += 1
"""
# if FN0 > TP0:
# print imgname
# problematic.append(imgname)
if print_steps: print(TP0, TN0, FP0, FN0)
# finalni vyhodnoceni
recall = float(TP) / (TP + FN) if TP + FN > 0 else 0
precision = float(TP) / (TP + FP) if TP + FP > 0 else 0
FPC = float(FP) / test_data.shape[0] if test_data.shape[0] > 0 else 0
# if orig_only:
# FPC = float(FP) / len([k for k in test_results_nms.keys() if not "AFFINE" in k])
print("[RESULT] Celkove vysledky pro " + str(test_data.shape[0]) +
" obrazku:")
print(" TP:", TP)
print(" TN:", TN)
print(" FP:", FP)
print(" FN:", FN)
print(" TPR:", recall)
print(" precision:", precision)
print(" FPC:", FPC)
results_to_save = {
"TP": TP,
"TN": TN,
"FP": FP,
"FN": FN,
"TPR": recall,
"recall": recall,
"precision": precision,
"FPC": FPC,
"problematic": problematic
}
return TN, FP, FN, TP, results_to_save
