def detect_morphological(
self, images: List[np.ndarray]
) -> Generator[List[List[List[List[int]]]], None, None]:
for img in images:
image_data = ImageData()
image_data.image = img.astype(float) / 255
gray = image_data.image.copy()
if np.sum(np.histogram(gray)[0][1:-2]) != 0:
gray = enhance(image_data.image)
binary = binarize(gray)
binarized = 1 - binary
morph = binary_erosion(binarized, structure=np.full((5, 1), 1))
morph = binary_dilation(morph, structure=np.full((5, 1), 1))
staffs = (binarized ^ morph)
image_data.staff_space_height, image_data.staff_line_height = vertical_runs(
binary)
image_data.binary_image = binary
self.callback.update_total_state()
image_data.horizontal_runs_img = calculate_horizontal_runs(
(1 - staffs), self.settings.horizontal_min_length)
self.callback.update_total_state()
yield self.detect_staff_lines(image_data)
self.callback.update_total_state()
self.callback.update_page_counter()
def create_data(image: np.ndarray, line_space_height: int) -> ImageData:
norm_img = image.astype(np.float32) / 255
binary_image = gauss_threshold(image.astype(np.uint8)) / 255
if line_space_height == 0:
staff_space_height, staff_line_height = vertical_runs(binary_image)
space_height = staff_space_height + staff_line_height
else:
space_height = line_space_height
staff_line_height = max(1, int(np.round(0.2 * line_space_height)))
staff_space_height = space_height - staff_line_height
image_data = ImageData(height=space_height,
image=norm_img,
staff_line_height=staff_line_height,
staff_space_height=staff_space_height,
binary_image=binary_image)
return image_data
def detect_fcn(
self, images: List[np.ndarray]
) -> Generator[List[List[List[List[int]]]], None, None]:
self.callback.set_total_pages(len(images))
create_data_partial = partial(
create_data, line_space_height=self.settings.line_space_height)
if len(images) <= 1:
data = list(map(create_data_partial, images))
else:
with multiprocessing.Pool(processes=self.settings.processes) as p:
data = [
v for v in tqdm.tqdm(p.imap(create_data_partial, images),
total=len(images))
]
for i, prob in enumerate(self.predictor.predict(data)):
prob = prob
if self.settings.model_use_argmax:
pred = np.argmax(prob, axis=-1)
else:
prob = prob[:, :, 1]
norm = prob / np.max(
prob) if self.settings.model_foreground_normalize else prob
pred = (norm > self.settings.model_foreground_threshold)
self.callback.update_total_state()
if data[i].staff_space_height is None or data[
i].staff_line_height is None:
data[i].staff_space_height, data[
i].staff_line_height = vertical_runs(data[i].binary_image)
data[i].horizontal_runs_img = calculate_horizontal_runs(
1 - pred, self.settings.horizontal_min_length)
data[i].pixel_classifier_prediction = prob
self.callback.update_total_state()
if self.settings.debug_model:
f, ax = plt.subplots(2, 3, sharex='all', sharey='all')
ax[0, 0].imshow(1 - prob, cmap='gray')
ax[0, 1].imshow(1 - pred, cmap='gray')
ax[0, 2].imshow(255 - data[i].horizontal_runs_img, cmap='gray')
ax[1, 0].imshow(data[i].image, cmap='gray')
plt.show()
yield self.detect_staff_lines(data[i])
self.callback.update_total_state()
self.callback.update_page_counter()
def detect_advanced(
self, images: List[np.ndarray]
) -> Generator[List[List[List[List[int]]]], None, None]:
create_data_partial = partial(
create_data, line_space_height=self.settings.line_space_height)
with multiprocessing.Pool(processes=self.settings.processes) as p:
data = [
v for v in tqdm.tqdm(p.imap(create_data_partial, images),
total=len(images))
]
self.callback.set_total_pages(len(images))
if self.text_predictor:
for i, pred in enumerate(
zip(self.predictor.predict(data),
self.text_predictor.predict(data))):
line_prediction = pred[0]
region_prediction = pred[1]
t_region = np.clip(pred[1], 0, 1) * 255
line_prediction[line_prediction > 0] = 255
region_prediction[region_prediction < 255] = 0
line_prediction_pruned = line_prediction * region_prediction * 255
self.callback.update_total_state()
if self.settings.debug:
x, y = plt.subplots(1, 3, sharex=True, sharey=True)
y[0].imshow(region_prediction)
y[1].imshow(line_prediction)
y[2].imshow(line_prediction_pruned)
plt.show()
self.callback.update_total_state()
data[i].staff_space_height, data[
i].staff_line_height = vertical_runs(
1 - line_prediction_pruned)
self.callback.update_total_state()
data[i].horizontal_runs_img = calculate_horizontal_runs(
(1 - (line_prediction_pruned / 255)),
self.settings.horizontal_min_length)
self.callback.update_total_state()
yield self.detect_staff_lines_rest(
data[i], get_text_borders(t_region - region_prediction))
self.callback.update_total_state()
self.callback.update_page_counter()
else:
for i, prob in enumerate(self.predictor.predict(data)):
background = prob[:, :, 0]
drawn_lines = prob[:, :, 1]
virtual_lines = prob[:, :, 2]
#prob = pred.probabilities[:, :, 1]
norm_vl = virtual_lines / np.max(virtual_lines) if self.settings.model_foreground_normalize else\
virtual_lines
pred_vl = (norm_vl > self.settings.model_foreground_threshold)
norm_dl = drawn_lines / np.max(
drawn_lines
) if self.settings.model_foreground_normalize else drawn_lines
pred_dl = (norm_dl > self.settings.model_foreground_threshold)
pred = pred_dl + pred_vl
self.callback.update_total_state()
data[i].staff_space_height, data[
i].staff_line_height = vertical_runs(1 - pred)
print(vertical_runs(1 - pred))
self.callback.update_total_state()
data[i].horizontal_runs_img = calculate_horizontal_runs(
(1 - pred), self.settings.horizontal_min_length)
plt.imshow(data[i].horizontal_runs_img)
plt.show()
self.callback.update_total_state()
data[i].image = images[i] / 255
binary = np.array(binarize(data[i].image), dtype='uint8')
text_borders = get_text_borders((1 - binary) * 255,
preprocess=True)
self.callback.update_total_state()
if self.settings.debug_model:
f, ax = plt.subplots(2, 3, sharex='all', sharey='all')
ax[0, 0].imshow(1 - prob, cmap='gray')
ax[0, 1].imshow(1 - pred, cmap='gray')
ax[0, 2].imshow(255 - data[i].horizontal_runs_img,
cmap='gray')
ax[1, 0].imshow(data[i].image, cmap='gray')
plt.show()
yield self.detect_staff_lines_rest(data[i], text_borders)
self.callback.update_total_state()
self.callback.update_page_counter()