def test_croptobbox():
X, Y = np.meshgrid(np.arange(16) - 8, np.arange(16) - 8)
ball = ((X**2 + Y**2) < 32).astype(np.uint8)
assert mahotas.croptobbox(ball).sum() == ball.sum()
assert mahotas.croptobbox(ball, border=2).sum() == ball.sum()
assert mahotas.croptobbox(ball, border=256).sum() == ball.sum()
assert mahotas.croptobbox(ball, border=256).size == ball.size
assert mahotas.croptobbox(ball.T).sum() == ball.sum()
assert mahotas.croptobbox(ball[::2]).sum() == ball[::2].sum()
assert mahotas.croptobbox(ball[::2].T).sum() == ball[::2].sum()
assert mahotas.croptobbox(ball.T, border=2).sum() == ball.sum()
assert mahotas.croptobbox(ball.T, border=256).sum() == ball.sum()
assert mahotas.croptobbox(ball.T, border=256).size == ball.size
def test_croptobbox():
X,Y = np.meshgrid(np.arange(16)-8,np.arange(16)-8)
ball = ((X**2+Y**2) < 32).astype(np.uint8)
assert mahotas.croptobbox(ball).sum() == ball.sum()
assert mahotas.croptobbox(ball,border=2).sum() == ball.sum()
assert mahotas.croptobbox(ball,border=256).sum() == ball.sum()
assert mahotas.croptobbox(ball,border=256).size == ball.size
assert mahotas.croptobbox(ball.T).sum() == ball.sum()
assert mahotas.croptobbox(ball[::2]).sum() == ball[::2].sum()
assert mahotas.croptobbox(ball[::2].T).sum() == ball[::2].sum()
assert mahotas.croptobbox(ball.T, border=2).sum() == ball.sum()
assert mahotas.croptobbox(ball.T, border=256).sum() == ball.sum()
assert mahotas.croptobbox(ball.T, border=256).size == ball.size
def view_labels(array, labels, crop=True, large=True, return_it=False):
if type(labels) != type(list()):
labels = [labels]
out = np.zeros(array.shape)
for l in labels:
l_arr = Util.threshold(array, l)
out[l_arr == 1] = out.max() + 1
if crop:
out = mh.croptobbox(out)
if large:
figsize = (10, 10)
else:
figsize = (3, 3)
if return_it:
return out
fig = plt.figure(figsize=figsize)
plt.imshow(out)
def view(array,
color=True,
large=False,
crop=False,
text=None,
no_axis=True,
file=''):
if large:
figsize = (10, 10)
else:
figsize = (3, 3)
fig = plt.figure(figsize=figsize)
if crop:
array = mh.croptobbox(array)
if text:
text = '\n\n\n' + str(text)
fig.text(0, 1, text)
if no_axis:
plt.axis('off')
if color:
plt.imshow(Util.colorize(array), picker=True)
else:
plt.imshow(array, cmap='gray', picker=True)
if file != '':
plt.savefig(file)
def _get_grid_img(self, crop_img, elem_contour):
cnt = Contour(elem_contour)
halfbox = cnt.shrinkbox(100).astype("f")
square = self._square_coordinates(cnt.side)
trans = cv2.getPerspectiveTransform(halfbox, square)
box = cv2.warpPerspective(crop_img, trans, (cnt.side, cnt.side))
_, thresh = cv2.threshold(box, 127, 255, 0)
binary = pm.binary(thresh)
box = binary.astype(int) * 255
box = pm.edgeoff(box)
box = mh.croptobbox(box)
return box
def view(array, color=True, large=False, crop=False, text=None):
if large:
figsize = (10, 10)
else:
figsize = (3, 3)
fig = plt.figure(figsize=figsize)
if crop:
array = mh.croptobbox(array)
if text:
text = '\n\n\n' + str(text)
fig.text(0, 1, text)
if color:
plt.imshow(Util.colorize(array), picker=True)
else:
plt.imshow(array, cmap='gray', picker=True)
def run_merge_test(cnn,
bbox=[0, 1024, 0, 1024],
data='gt',
slices=[70, 71, 72, 73, 74],
min_pixels=1000,
N=20,
oversampling=False,
keep_zeros=True,
verbose=False):
if data == 'rhoana':
print 'not implemented yet'
return [], []
print '-' * 80
print '-' * 80
print 'New Experiment:'
print ' Data:', data
print ' Slices:', slices
print ' No. splits for borders:', N
print ' Keep zeros in segmentation:', keep_zeros
# global_vis = []
global_vi_diffs = []
# global_surenesses = []
# global_merge_pairs = []
for s in slices:
if verbose:
print '-' * 80
print 'Working on slice', s
# load slice
input_image, input_prob, input_gold, input_rhoana = Util.read_section(
s, keep_zeros=keep_zeros)
# apply bbox
input_image = input_image[bbox[0]:bbox[1], bbox[2]:bbox[3]]
input_prob = input_prob[bbox[0]:bbox[1], bbox[2]:bbox[3]]
input_gold = input_gold[bbox[0]:bbox[1], bbox[2]:bbox[3]]
input_rhoana = input_rhoana[bbox[0]:bbox[1], bbox[2]:bbox[3]]
framed_gold = Util.frame_image(input_gold, shape=(200, 200))
hist = Util.get_histogram(framed_gold.astype(np.uint64))
labels = range(len(hist))
np.random.shuffle(labels)
slice_vi_diffs = []
for l in labels:
if l == 0:
continue
if len(framed_gold[framed_gold == l]) < min_pixels:
continue
neighbors = Util.grab_neighbors(framed_gold, l)
np.random.shuffle(neighbors)
good_neighbors = []
for n in neighbors:
if n == 0:
continue
if len(framed_gold[framed_gold == n]) < min_pixels:
continue
good_neighbors.append(n)
if len(good_neighbors) > 0:
for n in good_neighbors:
# print 'merging', l, n
before_merge_error = np.zeros(framed_gold.shape)
before_merge_error[framed_gold == l] = 1
before_merge_error[framed_gold == n] = 2
before_merge_error = mh.croptobbox(before_merge_error)
cropped_image, cropped_prob, cropped_segmentation, cropped_binary, bbox, real_border = Uglify.merge_label(
input_image,
input_prob,
framed_gold,
l,
n,
crop=True)
vi_before = Util.vi(
before_merge_error[10:-10,
10:-10].astype(np.uint8),
mh.croptobbox(cropped_binary)[10:-10,
10:-10].astype(
np.uint8))
# print 'VI after merge error:', vi_before
borders, best_border_image, result, result_no_border, results_no_border, predictions = Fixer.fix_single_merge(
cnn,
cropped_image,
cropped_prob,
cropped_binary,
real_border=real_border,
N=N,
erode=True,
oversampling=False)
if result_no_border.shape[0] == 0:
continue
if best_border_image.max() == 0:
# print 'no solution'
continue
# if before_merge_error.shape[0] != result_no_border.shape[0] or before_merge_error.shape[1] != result_no_border.shape[1]:
# result_no_border = np.resize(result_no_border, before_merge_error.shape)
# if before_merge_error.size != mh.croptobbox(r)
# compare_result = np.zeros(before_merge_error.shape, dtype=np.uint8)
# compare_result[:] = result_no_border[101:-101, 101:-101][0:before_merge_error.shape[0], 0:before_merge_error.shape[1]]
best_vi = np.inf
vi_diffs = []
# sorted_predictions = sorted(predictions)
for r in results_no_border:
if r.shape[0] == 0:
continue
r = r[100:-100, 100:-100]
result_no_border_center = (r.shape[0] / 2,
r.shape[1] / 2)
before_merge_center = (
before_merge_error.shape[0] / 2 - 10,
before_merge_error.shape[1] / 2 - 10)
r = r[result_no_border_center[0] -
before_merge_center[0]:
result_no_border_center[0] +
before_merge_center[0],
result_no_border_center[1] -
before_merge_center[1]:
result_no_border_center[1] +
before_merge_center[1]]
b = before_merge_error[result_no_border_center[0] -
before_merge_center[0]:
result_no_border_center[0] +
before_merge_center[0],
result_no_border_center[1] -
before_merge_center[1]:
result_no_border_center[1] +
before_merge_center[1]]
vi_after_fixing = Util.vi(b.astype(np.uint8),
r.astype(np.uint8))
vi_diffs.append(vi_before - vi_after_fixing)
# now we have vi_diffs for this one merge error
slice_vi_diffs.append((vi_diffs, predictions))
global_vi_diffs.append(slice_vi_diffs)
vi_correction_bins = [0, 0, 0, 0, 0]
bin_counts = [0, 0, 0, 0, 0]
for s in global_vi_diffs:
for merge_errors in s:
vi_diff_per_error = merge_errors[0]
predictions_per_error = merge_errors[1]
# sort by prediction
found_borders = sorted(zip(vi_diff_per_error,
predictions_per_error),
key=lambda x: x[1])
for i in range(5):
if len(found_borders) > i:
for j in range(i, len(found_borders)):
print i, j, len(vi_correction_bins), len(
found_borders)
vi_correction_bins[j] += found_borders[j][0]
bin_counts[j] += 1
for i in range(5):
vi_correction_bins[i] /= bin_counts[i]
return global_vi_diffs, vi_correction_bins
def test_croptobbox_3d():
YXZ = np.indices((32, 32, 64), float)
YXZ -= 8
Y, X, Z = YXZ
ball = ((X**2 + Y**2 + Z**2) < 64).astype(np.uint8)
assert np.sum(ball) == np.sum(mh.croptobbox(ball))
def croppBoundingBox(self):
# Returns a version of img cropped to the image’s bounding box.
# http://mahotas.readthedocs.io/en/latest/api.html#mahotas.croptobbox
return mh.croptobbox(self.imgArray, 0)
def test_croptobbox_3d():
YXZ = np.indices((32,32,64), float)
YXZ -= 8
Y,X,Z = YXZ
ball = ((X**2+Y**2+Z**2) < 64).astype(np.uint8)
assert np.sum(ball) == np.sum(mh.croptobbox(ball))
