avans06's picture
init commit.
8135b6a
import math
import os
import json
import sys
import time
import cv2 as cv
import numpy as np
from lib.panel import Panel
from lib.segment import Segment
from lib.debug import Debug
class NotAnImageException(Exception):
pass
class Page:
DEFAULT_MIN_PANEL_SIZE_RATIO = 1 / 10
def get_infos(self):
actual_gutters = self.actual_gutters()
return {
'filename': self.url if self.url else os.path.basename(self.filename),
'size': self.img_size,
'numbering': self.numbering,
'gutters': [actual_gutters['x'], actual_gutters['y']],
'license': self.license,
'panels': list(map(lambda p: p.to_xywh(), self.panels)),
'processing_time': self.processing_time
}
def __init__(
self,
filename,
numbering = None,
debug = False,
url = None,
min_panel_size_ratio = None,
panel_expansion = True
):
self.filename = filename
self.panels = []
self.segments = []
self.processing_time = None
t1 = time.time_ns()
self.img = cv.imread(filename)
if not isinstance(self.img, np.ndarray) or self.img.size == 0:
raise NotAnImageException(f"File {filename} is not an image")
self.numbering = numbering or "ltr"
if not (numbering in ['ltr', 'rtl']):
raise Exception('Fatal error, unknown numbering: ' + str(numbering))
self.small_panel_ratio = min_panel_size_ratio or Page.DEFAULT_MIN_PANEL_SIZE_RATIO
self.panel_expansion = panel_expansion
self.url = url
self.img_size = list(self.img.shape[:2])
self.img_size.reverse() # get a [width,height] list
Debug.contour_size = 3
# get license for this file
self.license = None
if os.path.isfile(filename + '.license'):
with open(filename + '.license', encoding = "utf8") as fh:
try:
self.license = json.load(fh)
except json.decoder.JSONDecodeError:
print(f"License file {filename+'.license'} is not a valid JSON file", file = sys.stderr)
sys.exit(1)
Debug.set_base_img(self.img)
Debug.add_step('Initial state', self.get_infos())
Debug.add_image('Input image')
self.gray = cv.cvtColor(self.img, cv.COLOR_BGR2GRAY)
Debug.add_image('Shades of gray', img = self.gray)
Debug.show_time("Shades of gray")
# https://docs.opencv.org/3.4/d2/d2c/tutorial_sobel_derivatives.html
ddepth = cv.CV_16S
grad_x = cv.Sobel(self.gray, ddepth, 1, 0, ksize = 3, scale = 1, delta = 0, borderType = cv.BORDER_DEFAULT)
# Gradient-Y
# grad_y = cv.Scharr(self.gray,ddepth,0,1)
grad_y = cv.Sobel(self.gray, ddepth, 0, 1, ksize = 3, scale = 1, delta = 0, borderType = cv.BORDER_DEFAULT)
abs_grad_x = cv.convertScaleAbs(grad_x)
abs_grad_y = cv.convertScaleAbs(grad_y)
self.sobel = cv.addWeighted(abs_grad_x, 0.5, abs_grad_y, 0.5, 0)
Debug.add_image('Sobel filter applied', img = self.sobel)
Debug.show_time("Sobel filter")
self.get_contours()
self.get_segments()
self.get_initial_panels()
self.group_small_panels()
self.split_panels()
self.exclude_small_panels()
self.merge_panels()
self.deoverlap_panels()
self.exclude_small_panels()
if self.panel_expansion:
self.panels.sort() # TODO: move this below before panels sort-fix, when panels expansion is smarter
self.expand_panels()
if len(self.panels) == 0:
self.panels.append(Panel(page = self, xywh = [0, 0, self.img_size[0], self.img_size[1]]))
self.group_big_panels()
self.fix_panels_numbering()
self.processing_time = int((time.time_ns() - t1) / 10**7) / 100
def get_contours(self):
# Black background: values above 100 will be black, the rest white
_, thresh = cv.threshold(self.sobel, 100, 255, cv.THRESH_BINARY)
Debug.show_time("Image threshhold")
self.contours, _ = cv.findContours(thresh, cv.RETR_EXTERNAL, cv.CHAIN_APPROX_SIMPLE)[-2:]
Debug.add_image("Thresholded image", img = thresh)
Debug.show_time("Get contours")
def get_segments(self):
self.segments = None
lsd = cv.createLineSegmentDetector(0)
dlines = lsd.detect(self.gray)
Debug.show_time("Detected segments")
min_dist = min(self.img_size) * self.small_panel_ratio
while self.segments is None or len(self.segments) > 500:
self.segments = []
if dlines is None or dlines[0] is None:
break
for dline in dlines[0]:
x0 = int(round(dline[0][0]))
y0 = int(round(dline[0][1]))
x1 = int(round(dline[0][2]))
y1 = int(round(dline[0][3]))
a = x0 - x1
b = y0 - y1
dist = math.sqrt(a**2 + b**2)
if dist >= min_dist:
self.segments.append(Segment([x0, y0], [x1, y1]))
min_dist *= 1.1
self.segments = Segment.union_all(self.segments)
Debug.draw_segments(self.segments, Debug.colours['green'])
Debug.add_image("Segment Detector")
Debug.show_time("Compiled segments")
# Get (square) panels out of initial contours
def get_initial_panels(self):
self.panels = []
for contour in self.contours:
arclength = cv.arcLength(contour, True)
epsilon = 0.001 * arclength
approx = cv.approxPolyDP(contour, epsilon, True)
panel = Panel(page = self, polygon = approx)
if panel.is_very_small():
continue
Debug.draw_contours([approx], Debug.colours['red'])
self.panels.append(panel)
Debug.add_image('Initial contours')
Debug.add_step('Panels from initial contours', self.get_infos())
# Group small panels that are close together, into bigger ones
def group_small_panels(self):
small_panels = list(filter(lambda p: p.is_small(), self.panels))
groups = {}
group_id = 0
for i, p1 in enumerate(small_panels):
for p2 in small_panels[i + 1:]:
if p1 == p2:
continue
if not p1.is_close(p2):
continue
if p1 not in groups and p2 not in groups:
group_id += 1
groups[p1] = group_id
groups[p2] = group_id
elif p1 in groups and p2 not in groups:
groups[p2] = groups[p1]
elif p2 in groups and p1 not in groups:
groups[p1] = groups[p2]
elif groups[p1] != groups[p2]:
# group group1 and group2 together
for p, id in groups.items():
if id == groups[p2]:
groups[p] = groups[p1]
grouped = {}
for k, v in groups.items():
grouped[v] = grouped.get(v, []) + [k]
for small_panels in grouped.values():
big_hull = cv.convexHull(np.concatenate(list(map(lambda p: p.polygon, small_panels))))
big_panel = Panel(page = self, polygon = big_hull, splittable = False)
self.panels.append(big_panel)
for p in small_panels:
self.panels.remove(p)
Debug.draw_contours(list(map(lambda p: p.polygon, small_panels)), Debug.colours['lightblue'])
Debug.draw_contours([big_panel.polygon], Debug.colours['red'])
if group_id > 0:
Debug.add_image('Group small panels')
Debug.add_step('Group small panels', self.get_infos())
# See if panels can be cut into several (two non-consecutive points are close)
def split_panels(self):
did_split = True
while did_split:
did_split = False
for p in sorted(self.panels, key = lambda p: p.area(), reverse = True):
split = p.split()
if split is not None:
did_split = True
self.panels.remove(p)
self.panels += split.subpanels
Debug.draw_contours(list(map(lambda n: n.polygon, split.subpanels)), Debug.colours['blue'])
Debug.draw_line(split.segment.a, split.segment.b, Debug.colours['red'])
break
if did_split:
Debug.add_image(
'Split contours (blue contours, red split-segment, gray polygon dots, purple nearby dots)'
)
Debug.add_step(f"Panels from split contours ({len(self.segments)} segments)", self.get_infos())
def exclude_small_panels(self):
self.panels = list(filter(lambda p: not p.is_small(), self.panels))
Debug.add_step('Exclude small panels', self.get_infos())
# Splitting polygons may result in panels slightly overlapping, de-overlap them
def deoverlap_panels(self):
for p1 in self.panels:
for p2 in self.panels:
if p1 == p2:
continue
opanel = p1.overlap_panel(p2)
if not opanel:
continue
if opanel.w() < opanel.h() and p1.r == opanel.r:
p1.r = opanel.x
p2.x = opanel.r
continue
if opanel.w() > opanel.h() and p1.b == opanel.b:
p1.b = opanel.y
p2.y = opanel.b
continue
Debug.add_step('Deoverlap panels', self.get_infos())
# Merge panels that shouldn't have been split (speech bubble diving into a panel)
def merge_panels(self):
panels_to_remove = []
for i, p1 in enumerate(self.panels):
for j, p2 in enumerate(self.panels[i + 1:]):
if p1.contains(p2):
panels_to_remove.append(p2)
p1 = p1.merge(p2)
elif p2.contains(p1):
panels_to_remove.append(p1)
p2 = p2.merge(p1)
for p in set(panels_to_remove):
self.panels.remove(p)
Debug.add_step('Merge panels', self.get_infos())
# Find out actual gutters between panels
def actual_gutters(self, func = min):
gutters_x = []
gutters_y = []
for p in self.panels:
left_panel = p.find_left_panel()
if left_panel:
gutters_x.append(p.x - left_panel.r)
top_panel = p.find_top_panel()
if top_panel:
gutters_y.append(p.y - top_panel.b)
if not gutters_x:
gutters_x = [1]
if not gutters_y:
gutters_y = [1]
return {'x': func(gutters_x), 'y': func(gutters_y), 'r': -func(gutters_x), 'b': -func(gutters_y)}
def max_gutter(self):
return max(self.actual_gutters().values())
# Expand panels to their neighbour's edge, or page boundaries
def expand_panels(self):
gutters = self.actual_gutters()
for p in self.panels:
for d in ['x', 'y', 'r', 'b']: # expand in all four directions
newcoord = -1
neighbour = p.find_neighbour_panel(d)
if neighbour:
# expand to that neighbour's edge (minus gutter)
newcoord = getattr(neighbour, {'x': 'r', 'r': 'x', 'y': 'b', 'b': 'y'}[d]) + gutters[d]
else:
# expand to the furthest known edge (frame around all panels)
min_panel = min(self.panels, key = lambda p: getattr(p, d)) if d in [
'x', 'y'
] else max(self.panels, key = lambda p: getattr(p, d))
newcoord = getattr(min_panel, d)
if newcoord != -1:
if d in ['r', 'b'] and newcoord > getattr(p, d) or d in ['x', 'y'] and newcoord < getattr(p, d):
setattr(p, d, newcoord)
Debug.add_step('Expand panels', self.get_infos())
# Fix panels simple sorting (issue #12)
def fix_panels_numbering(self):
changes = 1
while changes:
changes = 0
for i, p in enumerate(self.panels):
neighbours_before = [p.find_top_panel()]
neighbours_before += p.find_all_right_panels() if self.numbering == "rtl" else p.find_all_left_panels()
for neighbour in neighbours_before:
if neighbour is None:
continue
neighbour_pos = self.panels.index(neighbour)
if i < neighbour_pos:
changes += 1
self.panels.insert(neighbour_pos, self.panels.pop(i))
break
if changes > 0:
break # start a new whole loop with reordered panels
Debug.add_step('Numbering fixed', self.get_infos())
# group big panels together
def group_big_panels(self):
grouped = True
while grouped:
grouped = False
for i, p1 in enumerate(self.panels):
for p2 in self.panels[i + 1:]:
p3 = p1.group_with(p2)
other_panels = [p for p in self.panels if p not in [p1, p2]]
if p3.bumps_into(other_panels):
continue
# are there big segments in this panel?
segments = []
for s in self.segments:
if p3.contains_segment(s) and s.dist() > p3.diagonal().dist() / 5:
if s not in segments:
segments.append(s)
if len(segments) > 0: # maybe allow a small number of big segments here?
continue
self.panels.append(p3)
self.panels.remove(p1)
self.panels.remove(p2)
grouped = True
break
if grouped:
break
Debug.add_step('Group big panels', self.get_infos())