def thresholdCenterlines(nms, tLow=0.012, tHigh=0.12, bimodal=True):
""" Uses a continuity-preserving hysteresis thresholding to classify
centerlines.
Inputs:
nms -- Non-maxima suppressed singularity index response
Keyword Arguments:
bimodal -- true if the areas of rivers in the image are sufficiently
large that the distribution of ψ is bimodal
tLow -- lower threshold (automatically set if bimodal=True)
tHigh -- higher threshold (automatically set if bimodal=True)
Returns:
centerlines -- a binary matrix that indicates centerline locations
"""
if bimodal:
#Otsu's algorithm
nms = preprocess.double2im(nms, 'uint8')
tHigh,_ = cv2.threshold(nms, nms.min(), nms.max(), cv2.THRESH_OTSU)
tLow = tHigh * 0.1
strongCenterline = nms >= tHigh
centerlineCandidate = nms >= tLow
# Find connected components that has at least one strong centerline pixel
strel = np.ones((3, 3), dtype=bool)
cclabels, numcc = ndlabel(centerlineCandidate, strel)
sumstrong = ndsum(strongCenterline, cclabels, list(range(1, numcc+1)))
centerlines = np.hstack((0, sumstrong > 0)).astype('bool')
centerlines = centerlines[cclabels]
return centerlines
def thresholdCenterlines(nms, tLow=0.012, tHigh=0.12, bimodal=True):
""" Uses a continuity-preserving hysteresis thresholding to classify
centerlines.
Inputs:
nms -- Non-maxima suppressed singularity index response
Keyword Arguments:
bimodal -- true if the areas of rivers in the image are sufficiently
large that the distribution of ψ is bimodal
tLow -- lower threshold (automatically set if bimodal=True)
tHigh -- higher threshold (automatically set if bimodal=True)
Returns:
centerlines -- a binary matrix that indicates centerline locations
"""
if bimodal:
#Otsu's algorithm
nms = preprocess.double2im(nms, 'uint8')
tHigh,_ = cv2.threshold(nms, nms.min(), nms.max(), cv2.THRESH_OTSU)
tLow = tHigh * 0.1
strongCenterline = nms >= tHigh
centerlineCandidate = nms >= tLow
# Find connected components that has at least one strong centerline pixel
strel = np.ones((3, 3), dtype=bool)
cclabels, numcc = ndlabel(centerlineCandidate, strel)
sumstrong = ndsum(strongCenterline, cclabels, range(1, numcc+1))
centerlines = np.hstack((0, sumstrong > 0)).astype('bool')
centerlines = centerlines[cclabels]
return centerlines
def process_clusters1(self):
# throw away cluster sizes smaller than min_C and relabel
# gets the size of each cluster
clus_sz = ndsum(self.den_th,
self.clus_label,
index=np.arange(1, self.clus_nb + 1))
clus_rej = np.where(clus_sz < self.min_C)[0]
for i in [self.clus_slice[x] for x in clus_rej]:
self.den_th[i] = False
self.clus_label, self.clus_nb = label(self.den_th)
# cluster sizes in pixels
self.clus_sz = ndsum(self.den_th,
self.clus_label,
index=np.arange(1, self.clus_nb + 1)).astype(int)
self.keep_clus = range(-1, self.clus_nb + 1)
def getLargestMaskObject(mask):
'''Given a numpy array `mask' containing a binary mask, returns an equivalent array with only the largest mask object.'''
labeled, numfeatures = label(mask) # generate a feature label
sums = ndsum(mask, labeled, list(
range(numfeatures + 1))) # sum the pixels under each label
maxfeature = np.where(sums == max(
sums)) # choose the maximum sum whose index will be the label number
return mask * (labeled == maxfeature)
def cleanSegment(seg, fillHoles=True, keepLargest=True, minSize=0):
numUniqueVals = np.unique(seg).shape[0]
assert numUniqueVals > 1
foreground = seg != 0
fillable = binary_fill_holes(foreground) != foreground
seg1h = oneHot(seg, numUniqueVals)
seg1h[
...,
0] = 0 # zero out background so that argmax does not pick it as the first non-zero index
# clean each segmentation channel separately
for n in range(1, numUniqueVals):
segn = seg1h[..., n]
if np.unique(segn).shape[0] > 1:
labeled, numfeatures = label(
segn) # label each separate object with a different number
sums = ndsum(segn, labeled,
range(numfeatures +
1)) # sum the pixels under each label
# keep the largest feature or eliminate ones smaller than the minimum size
if keepLargest:
maxfeature = np.where(
sums == max(sums)
)[0] # choose the maximum sum whose index will be the label number
if len(maxfeature) > 0:
segn = (labeled == maxfeature[0]).astype(segn.dtype)
elif minSize > 0:
segn = segn.copy()
for i, s in enumerate(sums[1:]): # skip background
if s < minSize:
segn[labeled == (i + 1)] = 0
# fill holes, that is background areas surrounded by segmentation
if fillHoles:
# choose pixels that are background and are in areas filled by binary fill
holes = fillable * (binary_fill_holes(segn) != segn)
segn[holes] = 1
seg1h[..., n] = segn
seg = np.argmax(seg1h, seg1h.ndim - 1)
# eliminating smaller features from segmentation channels may create holes so fill these in
if fillHoles:
seg = greyFillHoles(seg)
return seg
def isolateLargestMask(mask):
'''Label the binary images in `mask' and return an image retaining only the largest.'''
labeled, numfeatures = label(
mask) # label each separate object with a different number
if numfeatures > 1: # if there's more than one object in the segmentation, keep only the largest as the best guess
sums = ndsum(mask, labeled,
range(numfeatures + 1)) # sum the pixels under each label
maxfeature = np.where(
sums == max(sums)
) # choose the maximum sum whose index will be the label number
mask = mask * (labeled == maxfeature
) # mask out the prediction under the largest label
return mask
def _calculate_focus_measure(self, src, operator, roi):
'''
see
IMPLEMENTATION OF A PASSIVE AUTOMATIC FOCUSING ALGORITHM
FOR DIGITAL STILL CAMERA
DOI 10.1109/30.468047
and
http://cybertron.cg.tu-berlin.de/pdci10/frankencam/#autofocus
'''
# need to resize to 640,480. this is the space the roi is in
# s = resize(grayspace(pychron), 640, 480)
src = grayspace(src)
v = crop(src, *roi)
di = dict(var=lambda x:variance(x),
laplace=lambda x: get_focus_measure(x, 'laplace'),
sobel=lambda x: ndsum(generic_gradient_magnitude(x, sobel, mode='nearest'))
)
func = di[operator]
return func(v)
def _calculate_focus_measure(self, src, operator, roi):
'''
see
IMPLEMENTATION OF A PASSIVE AUTOMATIC FOCUSING ALGORITHM
FOR DIGITAL STILL CAMERA
DOI 10.1109/30.468047
and
http://cybertron.cg.tu-berlin.de/pdci10/frankencam/#autofocus
'''
# need to resize to 640,480. this is the space the roi is in
# s = resize(grayspace(pychron), 640, 480)
src = grayspace(src)
v = crop(src, *roi)
di = dict(var=lambda x: variance(x),
laplace=lambda x: get_focus_measure(x, 'laplace'),
sobel=lambda x: ndsum(
generic_gradient_magnitude(x, sobel, mode='nearest')))
func = di[operator]
return func(v)
def closeSmallGaps(source_raster_path, output_dir):
# Read the image
with rasterio.open(source_raster_path) as src:
im = src.read(1)
image = im.copy()
logger.debug('Creating small cloud mask')
# Create a cloud mask
mask = im == FLAG_CLOUD
labels, count = label(mask)
areas = np.array(ndsum(mask, labels, np.arange(labels.max() + 1)))
# identify the clusters bigger than the ones that we want to interpolate
new_mask = areas > MAX_AREA_SIZE_TO_INTERPOLATE
big_clusters = new_mask[labels.ravel()].reshape(labels.shape)
# Reverse the new mask so we get the small cloud mask
small_cluster_mask = np.logical_xor(big_clusters, mask)
logger.debug('Small cloud mask created')
logger.debug('Applying inpainting algorithm')
# Apply OpenCV's inpaint algorithm to the image in so it fills the small clouds
image = opencvInpaint(image, small_cluster_mask)
logger.debug('Saving the image')
# Save the image in the same output directory
saveInterpolatedImage(image, source_raster_path, output_dir)