def normalized_scale(arr):
crss_denom = float(max(horizontal_transitions(arr)))
if crss_denom > 0:
crossings_val = (arr.shape[1]) / crss_denom
scale = 35.0 / crossings_val
else:
scale = 1.0
return scale
def combine_for_line_ind(self, line_info, lineind=None):
fli = [] # Final Line indices
flb = []
line_widths = []
line = line_info.lines_chars[lineind]
line.sort(key=lambda x: line_info.get_box(x)[0])
shapes = line_info.shapes
# line, line info, shapes
if line:
top = 1000000 # arbitrary high number
bottom = 0
for k in line:
j = line_info.get_box(k)
if j[1] < top:
top = j[1]
if j[1] + j[3] > bottom:
bottom = j[1] + j[3]
firstbox = line_info.get_box(line[0])
lastbox = line_info.get_box(line[-1])
whitespace = np.zeros(len(line), dtype=int)
for p, c in enumerate(line):
if p + 1 < len(line):
ab = line_info.get_box(c)
nab = line_info.get_box(line[p + 1])
ws_diff = nab[0] - (ab[0] + ab[2])
whitespace[p] = ws_diff
sum_whitespace = whitespace.sum()
ln_arr = shapes.img_arr[top:bottom, firstbox[0]:lastbox[0] +
lastbox[2]].copy()
try:
for inx in line_info.small_cc_lines_chars[lineind]:
box = line_info.get_box(inx)
ln_arr[box[1] - top:box[1] + box[3] - top, box[0] -
firstbox[0]:box[0] + box[2] - firstbox[0]] = 1
except:
pass
# Remove small contours when calculating scale
# for k in line_info.small_cc_lines_chars[lineind]:
# x, y, w, h = shapes.get_boxes()[k]
# x = x - firstbox[0]
# y = y - top
# ln_arr[y:y+h, x:x+w] = 1
crss_denom = float(max(horizontal_transitions(ln_arr)))
if crss_denom > 0:
crossings_val = (ln_arr.shape[1] - sum_whitespace) / crss_denom
scale = 35.0 / crossings_val
else:
scale = 1.0
self.transitions.append(scale)
else:
return []
line = iter(line)
# Initialize the current box and its attrs# BREAKWIDTH = 3.0
try:
cur_ind = [next(line)]
except StopIteration:
return []
# cb is current box, b is the next box
cb = line_info.get_box(cur_ind[0])
led, red, top, bottom, w, h = box_attrs(cb)
# Loop through box, combine and close along the way
for i in line:
b = line_info.get_box(i)
ledn, ren, topn, bottomn, wn, hn = box_attrs(b)
# If left edge of next box doesn't overlap cur box
# separate as 2 different chars
is_interior = ((ledn >= led and ren <= red) or
(ledn <= led and ren >= red)) and not bottom < topn
if not isinstance(i, str):
is_interior = line_info.shapes.hierarchy[0][i][
0] < 0 and line_info.shapes.hierarchy[0][i][
1] < 0 and line_info.shapes.hierarchy[0][i][
2] < 0 # i.e. it has no peers at its place in the tree... and no children
else:
is_interior = False # Its a string, meaning it is the result of a horizontal cut and likely not an interior
if ledn > red: # or (not is_interior and in_tsek_pop(shapes, wn,topn,top, bottomn, bs, i) and not hn > 1.5*shapes.tsek_mean
# ):
fli.append(cur_ind)
x, y, w, h = combine_many_boxes(
[line_info.get_box(j) for j in cur_ind])
flb.append((x, y, w, h))
self.widths.append(w * scale)
cur_ind = [i]
cb = b
led, red, top, bottom, w, h = box_attrs(cb)
# continue
else: # There is overlap
# one box is completely enveloped by the other
if is_interior:
# print 'enveloped'
cur_ind.append(i)
bxs = [line_info.get_box(j) for j in cur_ind]
bxs.append(b)
cb = combine_many_boxes(bxs)
led, red, top, bottom, w, h = box_attrs(cb)
elif ((float(min(wn, w)) - abs(
(red - ledn))) / float(min(wn, w))
) < self.hangoff: # amount hanging off end is 30 %
cur_ind.append(i)
bxs = [line_info.get_box(j) for j in cur_ind]
bxs.append(b)
cb = combine_many_boxes(bxs)
led, red, top, bottom, w, h = box_attrs(cb)
# The overlap is incidental / boxes are not related
else:
# print 'incidental overlap'
fli.append(cur_ind)
x, y, w, h = combine_many_boxes(
[line_info.get_box(j) for j in cur_ind])
flb.append((x, y, w, h))
self.widths.append(w * scale)
cur_ind = [i]
cb = b
led, red, top, bottom, w, h = box_attrs(cb)
fli.append(cur_ind)
x, y, w, h = combine_many_boxes(
[line_info.get_box(j) for j in cur_ind])
flb.append((x, y, w, h))
line_widths.append(w)
self.widths.append(w * scale)
if shapes.low_ink:
lib = self.line_info.low_ink_boxes[lineind]
low_ink_segmentation = {}
not_intr = []
for d, box in enumerate(flb):
# b = line_info.get_box(i)
led, red, top, bottom, w, h = box_attrs(box)
for p, bx in enumerate(lib):
ledn, ren, topn, bottomn, wn, hn = box_attrs(bx)
is_interior = (led >= ledn - 15
and red <= ren + 15) #and not bottom < topn
if is_interior:
### This attempts to remove noise
### that doesn't fall into blurred low ink
### box but does get combined according to normal
### combination rules
# for inx in fli[d]:
# tb = line_info.get_box(inx)
# if tb[1] >= topn and tb[1] + tb[3] <= bottomn:
# cur = low_ink_segmentation.get(p,[])
# # cur.extend(fli[d])
# cur.append(inx)
# low_ink_segmentation[p] = cur
cur = low_ink_segmentation.get(p, [])
cur.extend(fli[d])
low_ink_segmentation[p] = cur
break
else:
continue
else:
# print 'fail'
not_intr.append(fli[d])
#NO! do something
all_li_seg = low_ink_segmentation.values()
all_li_seg.extend(not_intr)
newfli = []
newflb = []
for j in all_li_seg:
newfli.append(j)
x, y, w, h = combine_many_boxes(
[line_info.get_box(i) for i in j])
newflb.append([x, y, w, h])
self.widths.append(w * scale)
fli = newfli
flb = newflb
fliflb = zip(fli, flb)
fliflb.sort(key=lambda x: x[1][0])
fli = [i[0] for i in fliflb]
flb = [i[1] for i in fliflb]
self.final_indices.append(fli)
self.final_boxes.append(flb)
self.line_width_means.append(mean(line_widths))
if not line:
print flb
def li_combine_for_line_ind(self, line_info, lineind=None):
fli = [] # Final Line indices
flb = []
line_widths = []
line = line_info.lines_chars[lineind]
lib = self.line_info.low_ink_boxes[lineind]
shapes = line_info.shapes
if line:
top = 1000000 # arbitrary high number
bottom = 0
for k in line:
j = line_info.get_box(k)
if j[1] < top:
top = j[1]
if j[1] + j[3] > bottom:
bottom = j[1] + j[3]
firstbox = line_info.get_box(line[0])
lastbox = line_info.get_box(line[-1])
whitespace = []
for p, c in enumerate(line):
if p + 1 < len(line):
ab = line_info.get_box(c)
nab = line_info.get_box(line[p + 1])
ws_diff = nab[0] - (ab[0] + ab[2])
# if ws_diff > 10*shapes.char_mean:
whitespace.append(ws_diff)
sum_whitespace = sum(whitespace)
ln_arr = shapes.img_arr[top:bottom, firstbox[0]:lastbox[0] +
lastbox[2]].copy()
crss_denom = float(max(horizontal_transitions(ln_arr)))
if crss_denom > 0:
crossings_val = (ln_arr.shape[1] - sum_whitespace) / crss_denom
scale = 35.0 / crossings_val
else:
scale = 1.0
self.transitions.append(scale)
else:
return []
line = iter(line)
low_ink_segmentation = {}
not_intr = []
for i in line:
b = line_info.get_box(i)
led, red, top, bottom, w, h = box_attrs(b)
for p, bx in enumerate(lib):
ledn, ren, topn, bottomn, wn, hn = box_attrs(bx)
is_interior = (led >= ledn - 5
and red <= ren + 5) #and not bottom < topn
if is_interior:
cur = low_ink_segmentation.get(p, [])
cur.append(i)
low_ink_segmentation[p] = cur
break
else:
continue
else:
not_intr.append([i])
all_li_seg = low_ink_segmentation.values()
all_li_seg.extend(not_intr)
for j in all_li_seg:
fli.append(j)
x, y, w, h = combine_many_boxes([line_info.get_box(i) for i in j])
flb.append([x, y, w, h])
self.widths.append(w * scale)
fliflb = zip(fli, flb)
fliflb.sort(key=lambda x: x[1][0])
fli = [i[0] for i in fliflb]
flb = [i[1] for i in fliflb]
self.final_indices.append(fli)
self.final_boxes.append(flb)