comicocr / modules /textdetector /panel_finder.py
fasdfsa's picture
init
f6f8d06
"""
Finds panel order for manga page.
>> python .\modules\textdetector\panel_finder.py <path-to-images>
"""
import json
import sys
from pathlib import Path
import cv2 as cv
import numpy as np
from PIL import Image, ImageDraw, ImageFont
from shapely import Polygon
from shapely.ops import nearest_points
KERNEL_SIZE = 7
BORDER_SIZE = 10
def panel_process_image(img: Image.Image):
"""Preprocesses an image to make it easier to find panels.
Args:
img: The image to preprocess.
Returns:
The preprocessed image.
"""
img_gray = cv.cvtColor(np.array(img), cv.COLOR_RGB2GRAY)
img_gray = cv.GaussianBlur(img_gray, (KERNEL_SIZE, KERNEL_SIZE), 0)
img_gray = cv.threshold(img_gray, 200, 255, cv.THRESH_BINARY)[1]
# Add black border to image, to help with finding contours
img_gray = cv.copyMakeBorder(
img_gray,
BORDER_SIZE,
BORDER_SIZE,
BORDER_SIZE,
BORDER_SIZE,
cv.BORDER_CONSTANT,
value=255,
)
# Invert image
img_gray = cv.bitwise_not(img_gray)
return img_gray
def remove_contained_contours(polygons):
"""Removes polygons from a list if any completely contain the other.
Args:
polygons: A list of polygons.
Returns:
A list of polygons with any contained polygons removed.
"""
# Create a new list to store the filtered polygons.
filtered_polygons = []
# Iterate over the polygons.
for polygon in polygons:
# Check if the polygon contains any of the other polygons.
contains = False
for other_polygon in polygons:
# Check if the polygon contains the other polygon and that the polygons
if np.array_equal(other_polygon, polygon):
continue
rect1 = cv.boundingRect(other_polygon)
rect2 = cv.boundingRect(polygon)
# Check if rect2 is completely within rect1
if (
rect2[0] >= rect1[0]
and rect2[1] >= rect1[1]
and rect2[0] + rect2[2] <= rect1[0] + rect1[2]
and rect2[1] + rect2[3] <= rect1[1] + rect1[3]
):
contains = True
break
# If the polygon does not contain any of the other polygons, add it to the
# filtered list.
if not contains:
filtered_polygons.append(polygon)
return filtered_polygons
def calc_panel_contours(im: Image.Image):
img_gray = panel_process_image(im)
contours_raw, hierarchy = cv.findContours(
img_gray, cv.RETR_EXTERNAL, cv.CHAIN_APPROX_SIMPLE
)
contours = contours_raw
min_area = 10000
contours = [i for i in contours if cv.contourArea(i) > min_area]
contours = [cv.convexHull(i) for i in contours]
contours = remove_contained_contours(contours)
# Remap the contours to the original image
contours = [i + np.array([[-BORDER_SIZE, -BORDER_SIZE]]) for i in contours]
# Sort the contours by their y-coordinate.
contours = order_panels(contours, img_gray)
return contours
def calc_panel_bboxes_xyxy(img: Image.Image):
contours = calc_panel_contours(img)
panel_bboxes = [cv.boundingRect(c) for c in contours]
panel_bboxes_xyxy = [xywh_to_xyxy(i) for i in panel_bboxes]
return panel_bboxes_xyxy
def draw_contours(im, contours):
"""Debugging, draws the contours on the image."""
colors = [
(255, 0, 0),
(0, 255, 0),
(0, 0, 255),
]
im_contour = np.array(im)
for i, contour in enumerate(range(len(contours))):
color = colors[i % len(colors)]
im_contour = cv.drawContours(im_contour, contours, i, color, 4, cv.LINE_AA)
# Draw a number at the top left of contour
x, y, _, _ = cv.boundingRect(contours[i])
cv.putText(
im_contour,
str(i),
(x + 50, y + 50),
cv.FONT_HERSHEY_SIMPLEX,
1,
color,
2,
cv.LINE_AA,
)
img = Image.fromarray(im_contour)
return img
def save_draw_contours(path: Path | str):
path = Path(path)
pth_out = path.parent / (path.stem + "-contours")
if not pth_out.exists():
pth_out.mkdir()
# Glob get all images in folder
pths = [i for i in path.iterdir() if i.suffix in [".png", ".jpg", ".jpeg"]]
for t in pths:
print(t)
im = Image.open(t)
contours = calc_panel_contours(im)
img_panels = draw_contours(im, contours)
f_name = t.stem + t.suffix
img_panels.save(pth_out / f_name)
def order_panels(contours, img_gray):
"""Orders the panels in a comic book page.
Args:
contours: A list of contours, where each contour is a list of points.
Returns:
A list of contours, where each contour is a list of points, ordered by
their vertical position.
"""
# Get the bounding boxes for each contour.
bounding_boxes = [cv.boundingRect(contour) for contour in contours]
# Generate groups of vertically overlapping bounding boxes.
groups_indices = generate_vertical_bounding_box_groups_indices(bounding_boxes)
c = []
for group in groups_indices:
# Reorder contours based on reverse z-order,
cs = [bounding_boxes[i] for i in group]
order_scores = order_read_direction_scores(cs)
# Sort the list based on the location score value
combined_list = list(zip(group, order_scores))
sorted_list = sorted(combined_list, key=lambda x: x[1], reverse=True)
c.extend(sorted_list)
ordered_contours = [contours[i[0]] for i in c]
return ordered_contours
def order_read_direction_scores(cs):
"""
Smaller means read first, larger means read last
"""
order_scores = [1 * (-i[1]) + i[0] * 1 for i in cs]
return order_scores
def generate_vertical_bounding_box_groups_indices(bounding_boxes):
"""Generates groups of vertically overlapping bounding boxes.
Args:
bounding_boxes: A list of bounding boxes, where each bounding box is a tuple
of (x, y, width, height).
Returns:
A list of groups, where each group is a list of bounding boxes that overlap
vertically.
"""
# Operate on indices Sort the bounding boxes by their y-coordinate.
bbox_inds = np.argsort([i[1] for i in bounding_boxes])
# generate groups of vertically overlapping bounding boxes
groups = [[bbox_inds[0]]]
for i in bbox_inds[1:]:
is_old_group = False
bbox = bounding_boxes[i]
start1 = bbox[1]
end1 = bbox[1] + bbox[3]
for n, group in enumerate(groups):
for ind in group:
_bbox = bounding_boxes[ind]
start2 = _bbox[1]
end2 = _bbox[1] + _bbox[3]
# Check for any partial overlapping
if check_overlap((start1, end1), (start2, end2)):
groups[n] = group + [i]
is_old_group = True
break
if is_old_group:
break
else:
groups.append([i])
return groups
def check_overlap(range1, range2):
# Check if range1 is before range2
if range1[1] < range2[0]:
return False
# Check if range1 is after range2
elif range1[0] > range2[1]:
return False
# If neither of the above conditions are met, the ranges must overlap
else:
return True
# Convert xyxy bounding boxes to shapely polygons
def polygon_from_xyxy(x, y, x2, y2):
return Polygon([(x, y), (x2, y), (x2, y2), (x, y2)])
def closest_text_to_panel_index(text_bboxes_xyxy, panel_bboxes_xyxy):
closest_boxes = []
# Iterate over each text bounding box
for t_index, text_box in enumerate(text_bboxes_xyxy):
# Initialize minimum distance to a large number
min_dist = float("inf")
# Initialize nearest box
# Convert text bounding box to Polygon
text_poly = polygon_from_xyxy(*text_box)
# Iterate over each panel bounding box
p_index = 0
for p_index, panel_box in enumerate(panel_bboxes_xyxy):
# Convert panel bounding box to Polygon
panel_poly = polygon_from_xyxy(*panel_box)
# Find the nearest points between the text and panel bounding boxes
nearest_pts = nearest_points(text_poly, panel_poly)
# Calculate the distance between the nearest points
dist = nearest_pts[0].distance(nearest_pts[1])
# If the distance is less than the minimum distance
if dist < min_dist:
# Update the minimum distance
min_dist = dist
# Update the nearest box
if not dist:
break
# Append the nearest box to the list of closest boxes
closest_boxes.append((p_index, t_index))
order_indices_dict = {i: [] for i in range(len(panel_bboxes_xyxy))}
for order_index in closest_boxes:
order_indices_dict[order_index[0]].append(order_index[1])
return order_indices_dict
def xywh_to_xyxy(xywh):
return [xywh[0], xywh[1], xywh[0] + xywh[2], xywh[1] + xywh[3]]
def xyxy_to_xywh(xyxy):
return [xyxy[0], xyxy[1], xyxy[2] - xyxy[0], xyxy[3] - xyxy[1]]
def reorder_boxes_indices(text_bboxes_xyxy, panel_bboxes_xyxy):
panel_text_order = closest_text_to_panel_index(text_bboxes_xyxy, panel_bboxes_xyxy)
box_orders = []
for i in range(len(panel_bboxes_xyxy)):
text_inds = panel_text_order[i]
orders = order_read_direction_scores(
[xyxy_to_xywh(i) for i in [text_bboxes_xyxy[i] for i in text_inds]],
)
# print(orders)
bbox_inds = np.argsort(orders)[::-1]
box_orders.extend([text_inds[i] for i in bbox_inds])
return box_orders
def draw_bboxes(img, text_bboxes_xyxy, panel_bboxes_xyxy):
image_ = img.copy()
# Create a drawing object
draw = ImageDraw.Draw(image_)
# Draw black boxes on the image
for i, box in enumerate(text_bboxes_xyxy):
# draw.rectangle(xywh_to_xyxy(box), fill="black")
draw.rectangle(box, outline="red")
draw.text(
box[:2],
str(i),
fill="red",
stroke_width=2,
font=ImageFont.truetype("arial.ttf", 50),
)
for i, box in enumerate(panel_bboxes_xyxy):
# draw.rectangle(xywh_to_xyxy(box), fill="black")
draw.rectangle(box, outline="blue")
draw.text(
box[:2],
str(i),
fill="blue",
stroke_width=2,
font=ImageFont.truetype("arial.ttf", 50),
)
# Show the image
return image_
def extract_text_info_from_ballons(data):
pages = data["pages"]
extracted_data = {
k1: [
{k: v for k, v in d.items() if k in ["text", "xyxy", "_bounding_rect"]}
for d in pages[k1]
]
for k1 in pages.keys()
}
return extracted_data
def text_bboxes_from_ballons(text_info):
text_bboxes_xyxy = [i["xyxy"] for i in text_info]
return text_bboxes_xyxy
def save_panel_text_order(path: Path | str):
path = Path(path)
path_json = path / (f"imgtrans_{path.stem}" + ".json")
pth_out = path.parent / (path.stem + "-panel-text-order")
if not pth_out.exists():
pth_out.mkdir()
# Glob get all images in folder
with open(path_json, encoding="utf8") as f:
data = json.load(f)
pages = data["pages"]
pages_keys = list(pages.keys())
for k in pages_keys:
page_info = pages[k]
text_bboxes = text_bboxes_from_ballons(page_info)
img = Image.open(path / k)
panel_bboxes = calc_panel_bboxes_xyxy(img)
text_reorderered_index = reorder_boxes_indices(text_bboxes, panel_bboxes)
text_bboxes = [text_bboxes[i] for i in text_reorderered_index]
img_out = draw_bboxes(img, text_bboxes, panel_bboxes)
img_out.save(pth_out / k)
def reorder_text_block_data(path: Path | str):
path = Path(path)
path_json = path / (f"imgtrans_{path.stem}" + ".json")
# Glob get all images in folder
with open(path_json, encoding="utf8") as f:
data = json.load(f)
pages = data["pages"]
pages_keys = list(pages.keys())
pages_reordered = {}
for k in pages_keys:
page_info = pages[k]
text_bboxes = text_bboxes_from_ballons(page_info)
img = Image.open(path / k)
panel_bboxes = calc_panel_bboxes_xyxy(img)
text_reorderered_index = reorder_boxes_indices(text_bboxes, panel_bboxes)
pages_reordered[k] = [page_info[i] for i in text_reorderered_index]
data["pages"] = pages_reordered
with open(path_json, 'w', encoding="utf8") as f:
json.dump(data, f)
if __name__ == "__main__":
save_draw_contours(sys.argv[1])
save_panel_text_order(sys.argv[1])