def teager(im):
im32 = im.astype(np.int32)
padded = np.pad(im32, (1, 1), 'edge')
return normalize_u8(3 * im32 ** 2 - padded[2:, 2:] * padded[:-2, :-2] / 2 \
- padded[2:, :-2] * padded[:-2, 2:] / 2 \
- padded[2:, 1:-1] * padded[:-2, 1:-1] \
- padded[1:-1, 2:] * padded[1:-1, :-2])
def remove_stroke_outliers(im, lines, k=1.0):
stroke_widths = fast_stroke_width(im)
if lib.debug:
lib.debug_imwrite('strokes.png',
lib.normalize_u8(stroke_widths.clip(0, 10)))
mask = np.zeros(im.shape, dtype=np.uint8)
for line in lines:
for letter in line:
sliced = letter.crop().apply(mask)
sliced |= letter.raster()
lib.debug_imwrite('letter_mask.png', -mask)
masked_strokes = stroke_widths.copy()
masked_strokes &= -mask
strokes_mean, strokes_std = masked_mean_std(masked_strokes, mask)
if lib.debug:
print('overall: mean:', strokes_mean, 'std:', strokes_std)
debug = cv2.cvtColor(im, cv2.COLOR_GRAY2RGB)
new_lines = []
for line in lines:
if len(line) <= 1: continue
good_letters = []
for letter in line:
crop = letter.crop()
if not crop.nonempty(): continue
raster = letter.raster()
sliced_strokes = crop.apply(stroke_widths).copy()
sliced_strokes &= lib.bool_to_u8(raster)
mean, std = masked_mean_std(sliced_strokes, raster)
if mean < strokes_mean - k * strokes_std:
if lib.debug:
print('skipping {:4d} {:4d} {:.03f} {:.03f}'.format(
letter.x,
letter.y,
mean,
std,
))
letter.box(debug, color=lib.RED)
else:
if lib.debug: letter.box(debug, color=lib.GREEN)
good_letters.append(letter)
if good_letters:
new_lines.append(TextLine(good_letters,
underlines=line.underlines))
lib.debug_imwrite("stroke_filter.png", debug)
return new_lines
def pca_gray(im):
assert len(im.shape) == 3
Lab = cv2.cvtColor(im, cv2.COLOR_BGR2Lab)
im_1d = Lab.reshape(im.shape[0] * im.shape[1], 3).astype(np.float32)
im_1d -= np.mean(im_1d)
U, S, V = np.linalg.svd(im_1d, full_matrices=False)
coeffs = V[0]
if coeffs[0] < 0:
coeffs = -coeffs
result = normalize_u8(np.tensordot(Lab, coeffs, axes=1))
lib.debug_imwrite('pca.png', result)
return result
def training_data(font_paths, font_size, W_h):
faces = [freetype.Face(font_path) for font_path in font_paths]
hi_res = np.concatenate([create_mosaic(face, font_size) for face in faces])
blurred_ims = [
cv2.GaussianBlur(hi_res, (0, 0), 7, 3),
cv2.GaussianBlur(hi_res, (0, 0), 3, 7),
]
blurred = np.concatenate(blurred_ims, axis=0)
hi_res_2 = np.tile(hi_res, (len(blurred_ims), 1))
lib.debug_imwrite('hi.png', hi_res_2)
lo_res = cv2.resize(blurred, (0, 0), None, 0.5, 0.5,
interpolation=cv2.INTER_AREA)
lib.debug_imwrite('lo.png', lo_res)
lo_res_hi, filtered_lo = features_lo(lo_res)
difference = hi_res_2.astype(np.float64) - lo_res_hi
lib.debug_imwrite('diff.png', lib.normalize_u8(difference))
# make sure we're on edges (in hi-res reference)
struct = cv2.getStructuringElement(cv2.MORPH_RECT, (3, 3))
gradient = cv2.morphologyEx(hi_res_2, cv2.MORPH_GRADIENT, struct)
gradient_means, _ = lib.mean_std(gradient, W_h)
patch_mask = gradient_means > np.percentile(gradient_means, 50)
# patch_centers should match others' shape.
step = 3
center_slice = slice(W_h // 2, -(W_h // 2) - 1, step)
patch_centers = patch_mask[center_slice, center_slice]
lo_patches = patches(filtered_lo, W_h, step)[patch_centers].transpose(0, 3, 1, 2)
hi_patches = patches(difference, W_h, step)[patch_centers]
t = lo_patches.shape[0]
lo_patches_vec = lo_patches.reshape(t, -1)
for i in range(lo_patches.shape[1]):
print_dict('lo_sq{}.png'.format(i),
lo_patches_vec[:, i * W_h * W_h:(i + 1) * W_h * W_h])
hi_patches_vec = hi_patches.reshape(t, W_h * W_h)
print_dict('hi_sq.png', hi_patches_vec)
# reduce dimensionality on lo-res patches with PCA.
pca = sklearn.decomposition.PCA(n_components=lo_patches_vec.shape[1] // 6)
Y_pca = pca.fit_transform(lo_patches_vec)
return Y_pca, hi_patches_vec, pca
def ntirogiannis2014(im):
lib.debug_prefix.append('ng2014')
debug_imwrite('input.png', im)
im_h, _ = im.shape
N, BG_prime = ng2014_normalize(im)
O = otsu(N)
debug_imwrite('O.png', O)
letters = algorithm.all_letters(O)
height_map = HeightMap(letters)
ratio_sum = 0
for h in range(1, height_map.max_height() + 1):
if len(height_map[h]) == 0: continue
ratio_sum += height_map.ratio_pixels(h) / height_map.ratio_components(
h)
if ratio_sum > 1:
break
min_height = h
if lib.debug: print('Accept components only >= height', h)
OP = O.copy()
for h in range(1, min_height):
for letter in height_map[h]:
sliced = letter.slice(OP)
np.place(sliced, letter.raster(), 255)
debug_imwrite('OP.png', OP)
strokes = fast_stroke_width(OP)
debug_imwrite('strokes.png', normalize_u8(strokes.clip(0, 10)))
SW = int(round(strokes.sum() / np.count_nonzero(strokes)))
if lib.debug: print('SW =', SW)
S = skeleton(OP)
debug_imwrite('S.png', S)
S_inv = ~S
# S_inv_32 = S_inv.astype(np.int32)
# FG_count = np.count_nonzero(S_inv)
FG_pos = im[S_inv.astype(bool)]
FG_avg = FG_pos.mean()
FG_std = FG_pos.std()
# FG = (S_inv & im).astype(np.int32)
# FG_avg = FG.sum() / float(FG_count)
# FG_std = np.sqrt(((S_inv_32 & (FG - FG_avg)) ** 2).sum() / float(FG_count))
if lib.debug: print('FG:', FG_avg, FG_std)
BG_avg = BG_prime.mean()
BG_std = BG_prime.std()
if lib.debug: print('BG:', BG_avg, BG_std)
if FG_avg + FG_std != 0:
C = -50 * np.log10((FG_avg + FG_std) / (BG_avg - BG_std))
k = -0.2 - 0.1 * C / 10
else: # This is the extreme case when the FG is 100% black, check the article explaination page before equation 5
C = -50 * np.log10((2.5) / (BG_avg - BG_std))
k = -0.2 - 0.1 * C / 10
if lib.debug: print('niblack:', C, k)
local = niblack(N, window_size=(2 * SW) | 1, k=k)
debug_imwrite('local.png', local)
local_CCs = algorithm.all_letters(local)
# NB: paper uses OP here, which results in neglecting all small components.
O_inv = ~O
O_inv_32 = O_inv.astype(np.int8,
copy=False).astype(np.int32).astype(np.uint32,
copy=False)
label_map_O_inv = O_inv_32 & local_CCs[0].label_map
CO_inv = np.zeros(im.shape, dtype=np.uint8)
for cc in local_CCs:
if np.count_nonzero(cc.slice(label_map_O_inv) == cc.label) / float(
cc.area()) >= C / 100:
CO_sliced = cc.slice(CO_inv)
np.place(CO_sliced, cc.raster(), 255)
CO = ~CO_inv
debug_imwrite('CO.png', CO)
CO_inv_dilated = cv2.dilate(CO_inv, rect33)
FB = ~(CO_inv | ((~O) & CO_inv_dilated))
debug_imwrite('FB.png', FB)
lib.debug_prefix.pop()
return FB
