404 lines
11 KiB
Python
404 lines
11 KiB
Python
import random
|
|
from typing import Tuple, Union, Iterable, List
|
|
|
|
import albumentations as A
|
|
import cv2 as cv
|
|
import numpy as np
|
|
import pytest
|
|
from PIL import Image, ImageOps
|
|
from PIL.Image import Transpose
|
|
|
|
Image_t = Union[Image.Image, np.ndarray]
|
|
#
|
|
# transform = A.Compose(
|
|
# [
|
|
# # geometric transforms
|
|
# A.HorizontalFlip(p=0.2),
|
|
# A.RandomRotate90(p=0.2),
|
|
# A.VerticalFlip(p=0.2),
|
|
# # brightness and contrast transforms
|
|
# A.OneOf(
|
|
# [
|
|
# A.RandomGamma(p=0.5),
|
|
# A.RandomBrightnessContrast(p=0.5),
|
|
# ],
|
|
# p=0.5,
|
|
# ),
|
|
# # noise transforms
|
|
# A.SomeOf(
|
|
# [
|
|
# A.Emboss(p=0.05),
|
|
# A.ImageCompression(p=0.05),
|
|
# A.PixelDropout(p=0.05),
|
|
# ],
|
|
# p=0.5,
|
|
# n=2,
|
|
# ),
|
|
# # color transforms
|
|
# A.SomeOf(
|
|
# [
|
|
# A.ColorJitter(p=1),
|
|
# A.RGBShift(p=1, r_shift_limit=0.1, g_shift_limit=0.1, b_shift_limit=0.1),
|
|
# A.ChannelShuffle(p=1),
|
|
# ],
|
|
# p=0.5,
|
|
# n=3, # 3 => all
|
|
# ),
|
|
# # blurring and sharpening transforms
|
|
# A.OneOf(
|
|
# [
|
|
# A.GaussianBlur(p=0.05),
|
|
# A.MotionBlur(p=0.05, blur_limit=21),
|
|
# A.Sharpen(p=0.05),
|
|
# ],
|
|
# p=0.5,
|
|
# ),
|
|
# # environmental transforms
|
|
# A.OneOf(
|
|
# [
|
|
# A.RandomRain(p=0.2, rain_type="drizzle"),
|
|
# A.RandomFog(p=0.2, fog_coef_upper=0.4),
|
|
# A.RandomSnow(p=0.2),
|
|
# ],
|
|
# p=0.5,
|
|
# ),
|
|
# ],
|
|
# p=0.5,
|
|
# )
|
|
from funcy import juxt, compose, identity, lflatten
|
|
|
|
from cv_analysis.locations import TEST_PAGE_TEXTURES_DIR
|
|
|
|
# transform = A.Compose(
|
|
# [
|
|
# # brightness and contrast transforms
|
|
# A.OneOf(
|
|
# [
|
|
# A.RandomGamma(p=0.2),
|
|
# A.RandomBrightnessContrast(p=0.2, brightness_limit=0.05, contrast_limit=0.05),
|
|
# ],
|
|
# p=0.5,
|
|
# ),
|
|
# # color transforms
|
|
# A.SomeOf(
|
|
# [
|
|
# A.ColorJitter(p=1),
|
|
# A.RGBShift(p=1, r_shift_limit=0.3, g_shift_limit=0.3, b_shift_limit=0.3),
|
|
# A.ChannelShuffle(p=1),
|
|
# ],
|
|
# p=1.0,
|
|
# n=3, # 3 => all
|
|
# ),
|
|
# # # blurring and sharpening transforms
|
|
# # A.OneOf(
|
|
# # [
|
|
# # A.GaussianBlur(p=0.05),
|
|
# # A.MotionBlur(p=0.05, blur_limit=21),
|
|
# # A.Sharpen(p=0.05),
|
|
# # ],
|
|
# # p=0.0,
|
|
# # ),
|
|
# ]
|
|
# )
|
|
from cv_analysis.utils.display import show_image
|
|
from cv_analysis.utils.drawing import draw_rectangles
|
|
from cv_analysis.utils.rectangle import Rectangle
|
|
|
|
transform = A.Compose(
|
|
[
|
|
# A.ColorJitter(p=1),
|
|
]
|
|
)
|
|
|
|
|
|
Color = Tuple[int, int, int]
|
|
|
|
|
|
@pytest.fixture(
|
|
params=[
|
|
# "rough_grain",
|
|
# "plain",
|
|
# "digital",
|
|
"crumpled",
|
|
]
|
|
)
|
|
def base_texture(request, size):
|
|
texture = Image.open(TEST_PAGE_TEXTURES_DIR / (request.param + ".jpg"))
|
|
texture = texture.resize(size)
|
|
return texture
|
|
|
|
|
|
@pytest.fixture(
|
|
params=[
|
|
"portrait",
|
|
# "landscape",
|
|
]
|
|
)
|
|
def orientation(request):
|
|
return request.param
|
|
|
|
|
|
@pytest.fixture(
|
|
params=[
|
|
# 30,
|
|
100,
|
|
]
|
|
)
|
|
def dpi(request):
|
|
return request.param
|
|
|
|
|
|
@pytest.fixture(
|
|
params=[
|
|
# "brown",
|
|
"sepia",
|
|
# "gray",
|
|
# "white",
|
|
# "light_red",
|
|
# "light_blue",
|
|
]
|
|
)
|
|
def color_name(request):
|
|
return request.param
|
|
|
|
|
|
@pytest.fixture(
|
|
params=[
|
|
# "smooth",
|
|
# "coarse",
|
|
"neutral",
|
|
]
|
|
)
|
|
def texture_name(request):
|
|
return request.param
|
|
|
|
|
|
@pytest.fixture(
|
|
params=[
|
|
# 30,
|
|
70,
|
|
# 150,
|
|
]
|
|
)
|
|
def color_intensity(request):
|
|
return request.param
|
|
|
|
|
|
def random_flip(image):
|
|
if random.choice([True, False]):
|
|
image = image.transpose(Transpose.FLIP_LEFT_RIGHT)
|
|
if random.choice([True, False]):
|
|
image = image.transpose(Transpose.FLIP_TOP_BOTTOM)
|
|
return image
|
|
|
|
|
|
@pytest.fixture
|
|
def color(color_name):
|
|
return {
|
|
"brown": "#7d6c5b",
|
|
"sepia": "#b8af88",
|
|
"gray": "#9c9c9c",
|
|
"white": "#ffffff",
|
|
"light_red": "#d68c8b",
|
|
"light_blue": "#8bd6d6",
|
|
}[color_name]
|
|
|
|
|
|
@pytest.fixture
|
|
def texture_fn(texture_name, size):
|
|
if texture_name == "smooth":
|
|
fn = blur
|
|
elif texture_name == "coarse":
|
|
fn = compose(overlay, juxt(blur, sharpen))
|
|
else:
|
|
fn = identity
|
|
|
|
return normalize_image_function(fn)
|
|
|
|
|
|
def blur(image: np.ndarray):
|
|
return cv.blur(image, (3, 3))
|
|
|
|
|
|
def normalize_image_format_to_array(image: Image_t):
|
|
if isinstance(image, Image.Image):
|
|
return np.array(image)
|
|
return image
|
|
|
|
|
|
def normalize_image_format_to_pil(image: Image_t):
|
|
if isinstance(image, np.ndarray):
|
|
return Image.fromarray(image)
|
|
return image
|
|
|
|
|
|
def normalize_image_function(func):
|
|
def inner(image):
|
|
image = normalize_image_format_to_array(image)
|
|
image = func(image)
|
|
image = normalize_image_format_to_pil(image)
|
|
return image
|
|
|
|
return inner
|
|
|
|
|
|
def sharpen(image: np.ndarray):
|
|
return cv.filter2D(image, -1, np.array([[-1, -1, -1], [-1, 6, -1], [-1, -1, -1]]))
|
|
|
|
|
|
def overlay(images, mode=np.sum):
|
|
assert mode in [np.sum, np.max]
|
|
images = np.stack(list(images))
|
|
image = mode(images, axis=0)
|
|
image = (image / image.max() * 255).astype(np.uint8)
|
|
return image
|
|
|
|
|
|
@pytest.fixture
|
|
def texture(base_texture, color, color_intensity):
|
|
color_image = Image.new("RGBA", base_texture.size, color)
|
|
color_image.putalpha(color_intensity)
|
|
texture = superimpose_texture_with_transparency(base_texture, color_image)
|
|
return texture
|
|
|
|
|
|
def tint_image(src, color="#FFFFFF"):
|
|
src.load()
|
|
r, g, b, alpha = src.split()
|
|
gray = ImageOps.grayscale(src)
|
|
result = ImageOps.colorize(gray, (0, 0, 0), color)
|
|
result.putalpha(alpha)
|
|
return result
|
|
|
|
|
|
def color_shift_array(image: np.ndarray, color: Color):
|
|
"""Creates a 3-tensor from a 2-tensor by stacking the 2-tensor three times weighted by the color tuple."""
|
|
assert image.ndim == 3
|
|
assert image.shape[-1] == 3
|
|
assert isinstance(color, tuple)
|
|
assert max(color) <= 255
|
|
assert image.max() <= 255
|
|
|
|
color = np.array(color)
|
|
weights = color / color.sum() / 10
|
|
assert max(weights) <= 1
|
|
|
|
colored = (image * weights).astype(np.uint8)
|
|
|
|
assert colored.shape == image.shape
|
|
|
|
return colored
|
|
|
|
|
|
@pytest.fixture
|
|
def size(dpi, orientation):
|
|
if orientation == "portrait":
|
|
size = (8.5 * dpi, 11 * dpi)
|
|
elif orientation == "landscape":
|
|
size = (11 * dpi, 8.5 * dpi)
|
|
else:
|
|
raise ValueError(f"Unknown orientation: {orientation}")
|
|
size = tuple(map(int, size))
|
|
return size
|
|
|
|
|
|
def superimpose_texture_with_transparency(page: Image, texture: Image) -> Image:
|
|
"""Superimposes a noise image with transparency onto a page image."""
|
|
assert page.mode == "RGB"
|
|
assert texture.mode == "RGBA"
|
|
assert page.size == texture.size
|
|
page.paste(texture, (0, 0), texture)
|
|
return page
|
|
|
|
|
|
@pytest.fixture
|
|
def blank_page(texture, texture_fn) -> np.ndarray:
|
|
"""Creates a blank page with a given orientation and dpi."""
|
|
page = random_flip(texture)
|
|
page = texture_fn(page)
|
|
content_box_generator = ContentBoxGenerator()
|
|
boxes = content_box_generator(page)
|
|
content_box_generator.draw_boxes(page, boxes)
|
|
return page
|
|
|
|
|
|
class ContentBoxGenerator:
|
|
def __init__(self):
|
|
self.left_margin_percentage = 0.05
|
|
self.right_margin_percentage = 0.05
|
|
self.top_margin_percentage = 0.1
|
|
self.bottom_margin_percentage = 0.1
|
|
|
|
self.margin_percentage = 0.005
|
|
self.max_depth = 3
|
|
self.initial_recursion_probability = 1
|
|
self.recursion_probability_decay = 0.1
|
|
|
|
def __call__(self, page: Image.Image) -> List[Rectangle]:
|
|
left_margin = int(page.width * self.left_margin_percentage)
|
|
right_margin = int(page.width * self.right_margin_percentage)
|
|
top_margin = int(page.height * self.top_margin_percentage)
|
|
bottom_margin = int(page.height * self.bottom_margin_percentage)
|
|
|
|
box = Rectangle(left_margin, top_margin, page.width - right_margin, page.height - bottom_margin)
|
|
boxes = lflatten(self.generate_content_boxes(box))
|
|
boxes = self.drop_small_boxes(boxes, *page.size)
|
|
return boxes
|
|
|
|
def draw_boxes(self, page: Image, boxes: Iterable[Rectangle]):
|
|
image = draw_rectangles(page, boxes, filled=False, annotate=True)
|
|
show_image(image)
|
|
|
|
def generate_content_boxes(self, box: Rectangle, depth=0):
|
|
if depth >= self.max_depth:
|
|
yield box
|
|
else:
|
|
child_boxes = self.generate_random_child_boxes(box)
|
|
if self.recurse(depth):
|
|
yield from (self.generate_content_boxes(b, depth + 1) for b in child_boxes)
|
|
else:
|
|
yield child_boxes
|
|
|
|
def generate_random_child_boxes(self, box: Rectangle) -> Tuple[Rectangle, Rectangle]:
|
|
axis = random.choice(["x", "y"])
|
|
|
|
edge_anchor_point, edge_length = (box.x1, box.width) if axis == "x" else (box.y1, box.height)
|
|
split_coordinate = random.uniform(0.3, 0.7) * edge_length + edge_anchor_point
|
|
child_boxes = self.get_child_boxes(box, split_coordinate, axis)
|
|
return child_boxes
|
|
|
|
def get_child_boxes(self, box: Rectangle, split_coordinate, axis) -> Tuple[Rectangle, Rectangle]:
|
|
def low(p):
|
|
return p * (1 + self.margin_percentage)
|
|
|
|
def high(p):
|
|
return p * (1 - self.margin_percentage)
|
|
|
|
if axis == "x":
|
|
return (
|
|
Rectangle(low(box.x1), low(box.y1), high(split_coordinate), high(box.y2)),
|
|
Rectangle(low(split_coordinate), low(box.y1), high(box.x2), high(box.y2)),
|
|
)
|
|
else:
|
|
return (
|
|
Rectangle(low(box.x1), low(box.y1), high(box.x2), high(split_coordinate)),
|
|
Rectangle(low(box.x1), low(split_coordinate), high(box.x2), high(box.y2)),
|
|
)
|
|
|
|
def recurse(self, depth):
|
|
return random.random() <= self.recursion_probability(depth)
|
|
|
|
def recursion_probability(self, depth):
|
|
return self.initial_recursion_probability * (1 - self.recursion_probability_decay) ** depth
|
|
|
|
def drop_small_boxes(
|
|
self,
|
|
boxes: Iterable[Rectangle],
|
|
page_width,
|
|
page_height,
|
|
min_percentage=0.13,
|
|
) -> List[Rectangle]:
|
|
min_width = page_width * min_percentage
|
|
min_height = page_height * min_percentage
|
|
return [b for b in boxes if b.width > min_width and b.height > min_height]
|