def text_line_segmentation_NN(image,
scale,
mask=None,
use_binary=False,
debug_image=None,
offset=(0, 0)):
h, w = image.shape[:2]
if debug_image is None:
debug_image = image
if len(debug_image.shape) < 3:
debug_image = cv2.cvtColor(image, cv2.COLOR_GRAY2BGR)
if len(image.shape) > 2:
image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
if use_binary:
image = morph.thresh_sauvola(image, k=0.15) * 255
if mask is not None:
image = (image + mask * 255).astype('uint8')
labels, _ = morph.label(image == 0)
objects = morph.find_objects(labels)
height_map = [
sl.height(o) for o in objects if sl.height(o) > 6
and sl.height(o) < 100 and sl.aspect_normalized(o) < 8
]
avg_h = max(np.nan_to_num(np.mean(height_map)), scale * 0.6)
block = Block(image, avg_h)
words = block.getWordBoundingBoxes()
lines = filter_and_merge_overlap_boxes(words,
max(avg_h, scale * 1.2) * 0.3)
lines = filter_and_merge_overlap_boxes(lines,
max(avg_h, scale * 1.2) * 0.3,
use_merge_same_line_only=True)
offset_x, offset_y = offset
# filter line by size
lines = [
l for l in lines if l[3] - l[1] > avg_h * 0.3 and l[3] - l[1] < avg_h *
2.5 and l[2] - l[0] > avg_h * 0.5
]
lines = [sl.pad_box(l, 0, (h, w)) for l in lines]
lines = [[
l[0] + offset_x, l[1] + offset_y, l[2] + offset_x, l[3] + offset_y
] for l in lines]
debug_image = block.paint(None)
return lines, debug_image
def remove_vlines(binary, gray, scale, maxsize=10):
labels, _ = morph.label(binary)
objects = morph.find_objects(labels)
for i, b in enumerate(objects):
if (sl.width(b) <= 20 and sl.height(b) > 200) or (sl.width(b) <= 45 and sl.height(b) > 500):
gray[b][labels[b] == i + 1] = 140
# gray[:,b[1].start:b[1].stop]=140
labels[b][labels[b] == i + 1] = 0
return array(labels != 0, "B")
def remove_vlines(binary,gray,scale,maxsize=10):
labels,_ = morph.label(binary)
objects = morph.find_objects(labels)
for i,b in enumerate(objects):
if (sl.width(b)<=20 and sl.height(b)>200) or (sl.width(b)<=45 and sl.height(b)>500):
gray[b][labels[b]==i+1] = 140
#gray[:,b[1].start:b[1].stop]=140
labels[b][labels[b]==i+1] = 0
return array(labels!=0, 'B')
def firstAnalyse(binary):
binaryary = morph.r_closing(binary.astype(bool), (1,1))
labels,_ = morph.label(binaryary)
objects = morph.find_objects(labels) ### <<<==== objects here
bysize = sorted(range(len(objects)), key=lambda k: sl.area(objects[k]))
# bysize = sorted(objects,key=sl.area)
scalemap = zeros(binaryary.shape)
smalldot = zeros(binaryary.shape, dtype=binary.dtype)
for i in bysize:
o = objects[i]
if amax(scalemap[o])>0:
# mask = where(labels[o] != (i+1),uint8(255),uint8(0))
# binary[o] = cv2.bitwise_and(binary[o],binary[o],mask=mask)
continue
scalemap[o] = sl.area(o)**0.5
scale = median(scalemap[(scalemap>3)&(scalemap<100)]) ### <<<==== scale here
for i,o in enumerate(objects):
if (sl.width(o) < scale/2) or (sl.height(o) < scale/2):
smalldot[o] = binary[o]
if sl.dim0(o) > 3*scale:
mask = where(labels[o] != (i+1),uint8(255),uint8(0))
binary[o] = cv2.bitwise_and(binary[o],binary[o],mask=mask)
continue
return objects, smalldot, scale
def simplefirstAnalyse(binary):
binaryary = morph.r_closing(binary.astype(bool), (1,1))
labels,_ = morph.label(binaryary)
objects = morph.find_objects(labels) ### <<<==== objects here
smalldot = zeros(binaryary.shape, dtype=binary.dtype)
scale = int(binary.shape[0]*0.7)
for i,o in enumerate(objects):
if (sl.width(o) < scale/2) or (sl.height(o) < scale/2):
smalldot[o] = binary[o]
if sl.dim0(o) > 3*scale:
mask = where(labels[o] != (i+1),uint8(255),uint8(0))
binary[o] = cv2.bitwise_and(binary[o],binary[o],mask=mask)
continue
return objects, smalldot, scale
def filter_junk_cc(binary, scale, maxsize):
junk_cc = np.zeros(binary.shape, dtype='B')
text_like = np.zeros(binary.shape, dtype='B')
labels, _ = morph.label(binary)
objects = morph.find_objects(labels)
for i, b in enumerate(objects):
if sl.width(b) > maxsize * scale or sl.area(b) > scale * scale * 8 or \
sl.aspect_normalized(b) > 8 or sl.min_dim(b) < scale * 0.35:
junk_cc[b][labels[b] == i + 1] = 1
else:
if sl.width(b) > 0.3 * scale and sl.height(b) > 0.3 * scale:
text_like[b][labels[b] == i + 1] = 1
return junk_cc, text_like
def detect_table(image, scale, maxsize=10, debug_path=None):
h, w = image.shape[:2]
if len(image.shape) > 2 and image.shape[2] >= 3:
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
else:
gray = image
# kernel = np.ones((5,5),np.uint8)
# gray = 255-cv2.morphologyEx(255-gray, cv2.MORPH_CLOSE, kernel)
binary = 1 - morph.thresh_sauvola(gray, k=0.05)
junk_cc, _ = filter_junk_cc(binary, scale, maxsize)
junk_cc = morph.r_closing(junk_cc, (5, 5))
print('calculating combine sep...')
combine_sep = compute_combine_seps(junk_cc, scale)
# using closing morphology to connect disconnected edges
close_thes = int(scale * 0.15)
closed_sep = morph.r_closing(combine_sep, (close_thes, close_thes))
if debug_path is not None:
cv2.imwrite(filename[:-4] + '_bin.png',
((1 - junk_cc) * 255).astype('uint8'))
cv2.imwrite(filename[:-4] + '_sep.png',
(closed_sep * 255).astype('uint8'))
labels, _ = morph.label(closed_sep)
objects = morph.find_objects(labels)
# result table list
boxes = []
for i, b in enumerate(objects):
if sl.width(b) > maxsize * scale or sl.area(
b) > scale * scale * 10 or (sl.aspect_normalized(b) > 6
and sl.max_dim(b) > scale * 1.5):
density = np.sum(combine_sep[b])
density = density / sl.area(b)
if (sl.area(b) > scale * scale * 10 and sl.min_dim(b) > scale * 1.0
and sl.max_dim(b) > scale * 8 and density < 0.4):
# calculate projection to determine table border
w = sl.width(b)
h = sl.height(b)
region = (labels[b] == i + 1).astype('uint8')
border_pad = max(w, h)
border_thres = scale * 2
proj_x = np.sum(region, axis=0)
proj_y = np.sum(region, axis=1)
proj_x[3:] += proj_x[:-3]
proj_y[3:] += proj_y[:-3]
sep_x = np.sort([j[0] for j in np.argwhere(proj_x > 0.75 * h)])
sep_y = np.sort([j[0] for j in np.argwhere(proj_y > 0.4 * w)])
# skip if sep count < 2
if len(sep_x) < 1 or len(sep_y) < 1: continue
border_left, border_right, border_top, border_bottom = None, None, None, None
if sep_x[0] < border_pad:
border_left = sep_x[0]
if sep_x[-1] > w - border_pad:
border_right = sep_x[-1]
if sep_y[0] < border_pad:
border_top = sep_y[0]
if sep_y[-1] > h - border_pad:
border_bottom = sep_y[-1]
# print_info(border_top, border_bottom, border_left, border_right)
if all([
j is not None for j in
[border_top, border_bottom, border_left, border_right]
]):
border_right = b[1].stop - b[1].start
boxes.append([
b[1].start + border_left, b[0].start + border_top,
b[1].start + border_right, b[0].start + border_bottom
])
# boxes.append(([b[1].start, b[0].start, b[1].stop, b[0].stop]))
return boxes