def kim2014(orig, O=None, split=False, n_points_w=None, n_tries=30):
lib.debug_imwrite('gray.png', binarize.grayscale(orig))
im = binarize.binarize(orig, algorithm=lambda im: binarize.sauvola_noisy(im, k=0.1))
global bw
bw = im
im_h, im_w = im.shape
AH, lines, _ = get_AH_lines(im)
if O is None:
O = np.array((im_w / 2.0, im_h / 2.0))
if split:
# Test if line start distribution is bimodal.
line_xs = np.array([line.left() for line in lines])
bimodal = line_xs.std() / im_w > 0.10
dual = bimodal and im_w > im_h
else:
dual = False
if dual:
print('Bimodal! Splitting page!')
pages = crop.split_lines(lines)
n_points_w = 1.2 * np.percentile(np.array([line.width() for line in lines]), 90)
n_points_w = max(n_points_w, 1800)
if lib.debug:
debug = cv2.cvtColor(bw, cv2.COLOR_GRAY2BGR)
for page in pages:
page_crop = Crop.from_lines(page).expand(0.005)
# print(page_crop)
page_crop.draw(debug)
lib.debug_imwrite('split.png', debug)
page_crops = [Crop.from_lines(page) for page in pages]
if len(page_crops) == 2:
[c0, c1] = page_crops
split_x = (c0.x1 + c1.x0) / 2
page_crops = [
c0.union(Crop(0, 0, split_x, im_h)),
c1.union(Crop(split_x, 0, im_w, im_h))
]
result = []
for i, (page, page_crop) in enumerate(zip(pages, page_crops)):
print('==== PAGE {} ===='.format(i))
lib.debug_prefix.append('page{}'.format(i))
page_image = page_crop.apply(orig)
page_bw = page_crop.apply(im)
page_AH, page_lines, _ = get_AH_lines(page_bw)
new_O = O - np.array((page_crop.x0, page_crop.y0))
lib.debug_imwrite('precrop.png', im)
lib.debug_imwrite('page.png', page_image)
bw = page_bw
dewarper = Kim2014(page_image, page_bw, page_lines, [page_lines],
new_O, page_AH, n_points_w)
result.append(dewarper.run_retry()[0])
lib.debug_prefix.pop()
return result
else:
lib.debug_prefix.append('page0')
dewarper = Kim2014(orig, im, lines, [lines], O, AH, n_points_w)
lib.debug_prefix.pop()
return dewarper.run_retry(n_tries=n_tries)
def process_image(original, dpi=None):
original_rot90 = original
for i in range(args.rotate / 90):
original_rot90 = np.rot90(original_rot90)
# original_rot90 = cv2.resize(original_rot90, (0, 0), None, 1.5, 1.5)
im_h, im_w = original_rot90.shape[:2]
# image height should be about 10 inches. round to 100
if not dpi:
dpi = int(round(im_h / 1100.0) * 100)
print('detected dpi:', dpi)
split = im_w > im_h # two pages
cropped_images = []
if args.dewarp:
lib.debug_prefix.append('dewarp')
dewarped_images = dewarp.kim2014(original_rot90)
for im in dewarped_images:
bw = binarize.binarize(im, algorithm=binarize.sauvola, resize=1.0)
lib.debug_prefix.append('crop')
_, [lines] = crop(im, bw, split=False)
lib.debug_prefix.pop()
c = Crop.from_lines(lines)
if c.nonempty():
cropped_images.append(Crop.from_whitespace(bw).apply(im))
lib.debug_prefix.pop()
else:
bw = binarize.binarize(original_rot90, algorithm=binarize.adaptive_otsu, resize=1.0)
debug_imwrite('thresholded.png', bw)
AH, line_sets = crop(original_rot90, bw, split=split)
for lines in line_sets:
c = Crop.from_lines(lines)
if c.nonempty():
lib.debug = False
bw_cropped = c.apply(bw)
orig_cropped = c.apply(original_rot90)
angle = algorithm.skew_angle(bw_cropped, original_rot90, AH, lines)
if not np.isfinite(angle): angle = 0.
rotated = algorithm.safe_rotate(orig_cropped, angle)
rotated_bw = binarize.binarize(rotated, algorithm=binarize.adaptive_otsu)
_, [new_lines] = crop(rotated, rotated_bw, split=False)
# dewarped = algorithm.fine_dewarp(rotated, new_lines)
# _, [new_lines] = crop(rotated, rotated_bw, split=False)
new_crop = Crop.union_all([line.crop() for line in new_lines])
if new_crop.nonempty():
# cropped = new_crop.apply(dewarped)
cropped = new_crop.apply(rotated)
cropped_images.append(cropped)
out_images = []
lib.debug_prefix.append('binarize')
for i, cropped in enumerate(cropped_images):
lib.debug_prefix.append('page{}'.format(i))
if lib.is_bw(original_rot90):
out_images.append(binarize.otsu(cropped))
else:
out_images.append(
binarize.ng2014_fallback(binarize.grayscale(cropped))
)
lib.debug_prefix.pop()
lib.debug_prefix.pop()
return dpi, out_images