def _find_blocks(self, array, trim_edges=False):
array = sp.clip(array, a_min=0, a_max=1)
temp = sp.pad(array, pad_width=1, mode='constant', constant_values=0)
end_pts = sp.where(sp.ediff1d(temp) == -1)[0] # Find 1->0 transitions
end_pts -= 1 # To adjust for 0 padding
seg_len = sp.cumsum(array)[end_pts]
seg_len[1:] = seg_len[1:] - seg_len[:-1]
start_pts = end_pts - seg_len + 1
a = dict()
a['start'] = start_pts
a['end'] = end_pts
a['length'] = seg_len
if trim_edges:
if (a['start'].size > 0) and (a['start'][0] == 0):
[a.update({item: a[item][1:]}) for item in a]
if (a['end'].size > 0) and (a['end'][-1] == sp.size(array)-1):
[a.update({item: a[item][:-1]}) for item in a]
return a
def _find_blocks(self, array, trim_edges=False):
array = sp.clip(array, a_min=0, a_max=1)
temp = sp.pad(array, pad_width=1, mode='constant', constant_values=0)
end_pts = sp.where(sp.ediff1d(temp) == -1)[0] # Find 1->0 transitions
end_pts -= 1 # To adjust for 0 padding
seg_len = sp.cumsum(array)[end_pts]
seg_len[1:] = seg_len[1:] - seg_len[:-1]
start_pts = end_pts - seg_len + 1
a = dict()
a['start'] = start_pts
a['end'] = end_pts
a['length'] = seg_len
if trim_edges:
if (a['start'].size > 0) and (a['start'][0] == 0):
[a.update({item: a[item][1:]}) for item in a]
if (a['end'].size > 0) and (a['end'][-1] == sp.size(array) - 1):
[a.update({item: a[item][:-1]}) for item in a]
return a
def segment(self, mol, threshold, end):
"""
takes a list whos values are some score of whether that specific coordinate is a linker.
this is typically used on the lists returned from partition but does not have to be.
All streches of elements whos threshold is above or below some value are then determined.
and a list of regions and there corresponding label are returned. It also ignores the last "end"
coordinates of mol.
regeions = [(0,end),(start,end),...,(start,end),(start,len(mol))]
labels = [1,0,...1,0] 1 for greater then threshold 0 for below.
"""
mask = mol > threshold
breaks = scipy.concatenate(
([0], scipy.argwhere(scipy.ediff1d(mask[:-end]))[:, 0] + 1,
[len(mol)]))
regions = scipy.array([(s, e)
for s, e in zip(breaks[:-1], breaks[1:])])
labels = scipy.array([int(mask[s]) for s, e in regions])
return (regions, labels)
def scan_receipt(img, clf):
angle_to_rotate = detect_skew(img, quality='medium')
print 'angle_to_rotate', angle_to_rotate
img = rotate(img, angle_to_rotate, order=5, cval=1.0)
thresh = threshold_otsu(img)
img = img < thresh
# plt.imshow(img, interpolation='nearest', cmap='gray')
# plt.axis('image')
# plt.xticks([])
# plt.yticks([])
# plt.show()
# extract individual lines
num_blacks = sp.sum(img, axis=1)
has_characters = num_blacks > 0.01 * img.shape[1]
line_diffs = sp.ediff1d(has_characters.astype(sp.int32))
line_starts = sp.nonzero(line_diffs == 1)[0]
line_ends = sp.nonzero(line_diffs == -1)[0] + 2
line_heights = line_ends - line_starts
normal_line_height = sp.median(line_heights)
lines = []
for start, end, height in zip(line_starts, line_ends, line_heights):
if height < normal_line_height * 1.2 and height > 5:
lines.append(img[start:end])
receipt_content = ''
# extract individual characters from lines
for line in lines:
line_content = ''
num_blacks = sp.sum(line, axis=0)
has_character = num_blacks > 0
col_diffs = sp.ediff1d(has_character.astype(sp.int32))
col_starts = sp.nonzero(col_diffs == 1)[0]
col_ends = sp.nonzero(col_diffs == -1)[0] + 2
char_widths = col_ends - col_starts
normal_char_width = sp.median(char_widths)
min_char_width = normal_char_width * 0.8
max_char_width = normal_char_width * 1.2
# account for cases when 2 chars are merged, or if a single char is separated
chars = []
i = 0
while i < len(col_starts):
start = col_starts[i]
end = col_ends[i]
width = end - start
if width < min_char_width:
while i + 1 < len(col_starts):
expanded_width = col_ends[i+1] - start
if expanded_width < min_char_width:
i += 1
elif expanded_width > max_char_width:
break
elif col_starts[i+1] - end >= 3:
break
else:
i += 1
break
chars.append(line[:, start:col_ends[i]])
elif width >= normal_char_width * 2:
chars.append(line[:, start:start+width//2-1])
chars.append(line[:, start+width//2+1:end])
else:
chars.append(line[:, start:end])
i += 1
# pad each char to a square and resize it to 24x24
padded_chars = []
for char in chars:
height, width = char.shape
if height > width:
diff = height - width
pad_left = diff // 2
padded_char = sp.zeros((height, height))
padded_char[:, pad_left:pad_left+width] = char
elif width > height:
diff = width - height
pad_top = diff // 2
padded_char = sp.zeros((width, width))
padded_char[pad_top:pad_top+height, :] = char
final_char = imresize(padded_char, (25, 25), 'bicubic')
padded_chars.append(final_char)
for char in padded_chars:
plt.imshow(char, interpolation='nearest', cmap='gray')
plt.axis('image')
plt.xticks([])
plt.yticks([])
plt.show()
