def test_2D_indices_col(self):
"""
The returned 2D index should be the same as control.
"""
control = numpy.where(self.array_2D == 66)
wh = numpy.where(self.array_1D_1 == 66)
ind2D = array_indices(array=self.array_2D, index=wh[0])
n = len(control[1])
self.assertEqual((control[1] == ind2D[1]).sum(), n)
def test_2D_indices_col(self):
"""
The returned 2D index should be the same as control.
"""
control = numpy.where(self.array_2D == 66)
wh = numpy.where(self.array_1D_1 == 66)
ind2D = array_indices(array=self.array_2D, index=wh[0])
n = len(control[1])
self.assertEqual((control[1] == ind2D[1]).sum(), n)
def test_dimensions(self):
"""
Test that the dimensions keyword works and yields the same result as the control.
"""
control = numpy.where(self.array_3D == 66)
control_sum = numpy.sum(self.array_3D[control])
wh = numpy.where(self.array_1D_2 == 66)
ind3D = array_indices(array=self.array_3D.shape, index=wh[0], dimensions=True)
test_sum = numpy.sum(self.array_3D[ind3D])
self.assertEqual(control_sum, test_sum)
def test_dimensions(self):
"""
Test that the dimensions keyword works and yields the same result as the control.
"""
control = numpy.where(self.array_3D == 66)
control_sum = numpy.sum(self.array_3D[control])
wh = numpy.where(self.array_1D_2 == 66)
ind3D = array_indices(array=self.array_3D.shape,
index=wh[0],
dimensions=True)
test_sum = numpy.sum(self.array_3D[ind3D])
self.assertEqual(control_sum, test_sum)
def getSegmentLocations(self, segmentID=1):
"""
Retrieve the pixel locations corresponding to a segmentID.
"""
ri = self.ri
i = segmentID
if ri[i+1] > ri[i]:
idx = ri[ri[i]:ri[i+1]]
idx = array_indices(self.dims, idx, dimensions=True)
else:
idx = (numpy.array([]), numpy.array([]))
return idx
def obj_centroid(array):
"""
Calculates centroids per object.
"""
dims = array.shape
h = histogram(array.flatten(), Min=1, reverse_indices='ri')
hist = h['histogram']
ri = h['ri']
cent = []
for i in numpy.arange(hist.shape[0]):
if (hist[i] == 0):
continue
idx = numpy.array(array_indices(dims, ri[ri[i]:ri[i+1]], dimensions=True))
cent_i = numpy.mean(idx, axis=1)
cent.append(cent_i)
return cent
def obj_rectangularity(array):
"""
Calculates rectangularity per object.
"""
dims = array.shape
h = histogram(array.flatten(), Min=1, reverse_indices='ri')
hist = h['histogram']
ri = h['ri']
rect = []
for i in numpy.arange(hist.shape[0]):
if (hist[i] == 0):
continue
idx = numpy.array(array_indices(dims, ri[ri[i]:ri[i+1]], dimensions=True))
min_yx = numpy.min(idx, axis=1)
max_yx = numpy.max(idx, axis=1)
diff = max_yx - min_yx + 1 # Add one to account for zero based index
bbox_area = numpy.prod(diff)
rect.append(hist[i] / bbox_area)
return rect
# Create the mask via the thresholds
mask = (array >= lower) & (array <= upper)
# The label function segments the image into contiguous blobs
label_array, num_labels = ndimage.label(mask, structure=s)
# Find the labels associated with the ROI
labels = label_array[ROIPixels]
mx_lab = numpy.max(labels)
# Find unique labels, excluding zero (background)
ulabels = (numpy.unique(labels[labels > 0])
).tolist() # Convert to list; Makes for neater indexing
# Generate a histogram to find the label locations
h = histogram(label_array.flatten(),
min=0,
max=mx_lab,
reverse_indices='ri')
hist = h['histogram']
ri = h['ri']
for lab in ulabels:
if hist[lab] == 0:
continue
idx.extend(ri[ri[lab]:ri[lab + 1]])
idx = numpy.array(idx)
idx = array_indices(dims, idx, dimensions=True)
return idx
def obj_get_boundary_method1(labelled_array, fill_holes=True):
"""
Get the pixels that mark the object boundary/perimeter.
Method 1.
8 neighbourhood chain code
5 6 7
4 . 0
3 2 1
4 neighbourhood chain code
. 3 .
2 . 0
. 1 .
"""
dims = labelled_array.shape
rows = dims[0]
cols = dims[1]
if fill_holes:
orig_binary = (labelled_array > 0).astype('uint8')
fill = obj_fill_holes(labelled_array)
s = [[1,1,1],[1,1,1],[1,1,1]]
labelled_array, nlabels = ndimage.label(fill, structure=s)
# We'll opt for the perimeter co-ordinates to be ordered in a clockwise fashion. GIS convention???
pix_directions = numpy.array([[ 0, 1],
[ 1, 1],
[ 1, 0],
[ 1,-1],
[ 0,-1],
[-1,-1],
[-1, 0],
[-1, 1]])
# Set up the distances as we traverse across a pixel
diag = numpy.sqrt(2.0) # NumPy will return a float64, but just in case future versions change....
pix_distances = {0 : 1.0,
1 : diag,
2 : 1.0,
3 : diag,
4 : 1.0,
5 : diag,
6 : 1.0,
7, diag
}
# Determine the co-ordinates (indices) of each segement
# The first index of each segment will be used to define the start and end of a boundary/perimeter
h = histogram(labelled_array.flatten(), Min=1, reverse_indices='ri')
hist = h['histogram']
ri = h['ri']
nlabels = hist.shape[0]
seg_start_idxs = numpy.zeros(nlabels, dtype='int')
# Boundary or perimeter ?? Will go with perimeter, akin to a method implement earlier which uses a
# convolution operator to determine perimeter length.
# Obtain the start indices of each segment/object
for i in numpy.arange(nlabels):
#if (hist[i] == 0): # The labeled array should be consecutive
# continue
seg_start_idxs[i] = ri[ri[i]:ri[i+1]][0] # Return the first index
# Convert the 1D indices to 2D indices used by NumPy
seg_start_idxs = array_indices(dims, seg_start_idxs, dimensions=True)
# Lots to figure out here. Dealing with 'from' and 'too' directions can make things confusing
# Keep track of the direction we last travelled, that way we can start at the next clockwise direction
# For single pixel objects or 'islands' use the histogram value to skip the follow boundary/search routine
"""
The memory of the order of the array should be 'C-Style': column, column, column, then next row.
eg a 2x5 array
0, 1, 2, 3, 4
5, 6, 7, 8, 9
Therefore the first index will only have labels to the right and below (below left, below right).
eg an object with a label ID of 6
0, 0, 6, 6, 6
6, 6, 6, 6, 6
As such, the first direction travelled will be to the right '0' in the freeman chain, and the first index will be the
first boundary co-ordinate and will be the final co-ordinate to close the boundary thereby creating a polygon.
As for linear features...
"""
# Probably deal with these inside the boundary tracking routine
# Let the tracking routine handle to/from and just return the final result
to_ = 0
from_ = 4
perimeter_info = {}
for i in range(hist.shape[0]):
if hist[i] == 0:
continue
if hist[i] == 1:
# What is the perimeter of a single pixel, 0.0, 4.0???
#perimeter_co_ords[i] = seg_start_idxs[i[0],i[1]] # Still need to design the function and how to return the result
continue
idx = (seg_start_idxs[0][i], seg_start_idxs[1][i])
label = i + 1
perimeter_info[i] = track_object_boundary(labelled_array, start_index=idx, label_id=label)
# Might need to format the perimeter_info dictionary before returning, ie turn the co-ords into numpy arrays.
# Or even into a polygon object using the shapely library???
# Using shapely might be easier to report geometrical attributes
# Still need to deal with holes within an object as ENVI does. They will increase an objects perimeter length.
# SciPy have a binary_fill_holes function. label the filled array, then get the indices, and then retrive only those indices
# for each object that are 0 in the original array.
# That might be one way to do it, which means re-writing the above function....ughhh :)
#!!!!This isn't the correct place for the handling of object holes, but just get the rough structure out!!!
if fill_holes:
for i in range(hist.shape[0]):
if hist[i] == 0:
continue
if hist[i] == 1:
# What is the perimeter of a single pixel, 0.0, 4.0???
continue
idx = (seg_start_idxs[0][i], seg_start_idxs[1][i])
label = i + 1
perimeter_info[i] = track_object_boundary(labelled_array, start_index=idx, label_id=label)
# Can we trust that the labelling of the filled and unfilled arrays will give the same object index??
# If we can we could use the area differences to determine if there are holes and only go through the
# hole perimeter tracking if needed.
single_object = numpy.zeros((rows*cols), dtype='uint8')
single_object[ri[ri[i]:ri[i+1]]] = 1
holes = numexpr.evaluate("(single_object - orig_binary) == 1")
labs, nlabs = ndimage.label(holes, s)
h_holes = histogram(labs, Min=1, reverse_indices='ri')
hist_holes = h_holes['histogram']
ri_h = h_holes['ri']
seg_holes_start_idxs = numpy.zeros(nlabs, dtype='int')
for j in numpy.arange(nlabels):
#if (hist[i] == 0): # The labeled array should be consecutive
# continue
seg_holes_start_idxs[j] = ri_h[ri_h[j]:ri_h[j+1]][0] # Return the first index
# Convert the 1D indices to 2D indices used by NumPy
seg_holes_start_idxs = array_indices(dims, seg_holes_start_idxs, dimensions=True)
for k in range(hist_holes.shape):
if hist_holes[k] == 0:
continue
if hist[i] == 1:
# What is the perimeter of a single pixel, 0.0, 4.0???
continue
idx = (seg_holes_start_idxs[0][k], seg_holes_start_idxs[1][k])
holes_label = k + 1
holes_result = track_object_boundary(labs, start_index=idx, label_id=holes_label)
perimeter_info[i]['Holes'] = holes_result['Vertices']
perimeter_info[i]['Perimeter_Length'] += perimeter_info['Perimeter_Length']
return perimeter_info
