controlled_downscale.py

import argparse
import sys

from PIL import Image
from typing import List, Optional, Tuple


Pos = Tuple[int, int]
Dim = Tuple[int, int]


class Box:
    def __init__(self, min: Pos, max: Pos) -> None:
        self._min = min
        self._max = max

    # inclusive
    def min(self) -> Tuple[int, int]:
        return self._min
    
    # inclusive
    def max(self) -> Tuple[int, int]:
        return self._max

    def width(self) -> int:
        return self._max[0] - self._min[0] + 1
    
    def height(self) -> int:
        return self._max[1] - self._min[1] + 1

    def dimensions(self) -> Tuple[int, int]:
        return (self.width(), self.height())

    # (left, upper, right, lower)
    def as_tuple(self) -> Tuple[int, int, int, int]:
        return (self._min[0], self._min[1], self._max[0], self._max[1])


class DownBox(Box):
    def __init__(self, min: Pos, max: Pos, down_pos: Pos) -> None:
        super().__init__(min, max)
        self._down_pos = down_pos

    def down_pos(self) -> Tuple[int, int]:
        return self._down_pos


class ExtractedBoxes:
    def __init__(self, boxes: List[DownBox]) -> None:
        self._boxes = boxes

    def boxes(self) -> List[DownBox]:
        return self._boxes

    def down_dimensions(self) -> Dim:
        if len(self._boxes) == 0:
            return (0, 0)
        back = self._boxes[-1]
        down = back.down_pos()
        return (down[0] + 1, down[1] + 1)

    def full_dimensions(self) -> Dim:
        if len(self._boxes) == 0:
            return (0, 0)
        back = self._boxes[-1]
        max = back.max()
        return (max[0] + 1, max[1] + 1)

    def to_colored_checkers(self, *, full=True) -> Image.Image:
        if full:
            width, height = self.full_dimensions()
        else:
            width, height = self.down_dimensions()
        if width == 0 or height == 0:
            return Image.new("RGB", (0, 0))
        image = Image.new("RGB", (width, height))
        colors = [
            (255, 255, 255),
            (0, 0, 0),
            (255, 0, 0),
            (255, 127, 0),
            (255, 255, 0),
            (0, 255, 0),
            (0, 0, 255),
            (75, 0, 130),
            (148, 0, 211),
            (255, 0, 255),
        ]
        colorsMax = len(colors)
        currColor = 0
        for box in self._boxes:
            color = colors[currColor]
            currColor = (currColor + 1) % colorsMax
            if full:
                dim = box.dimensions()
                pos = box.min()
            else:
                dim = (1, 1)
                pos = box.down_pos()
            subImage = Image.new("RGB", dim, color)
            image.paste(subImage, pos)
        return image


def average_box_dimensions(boxes: List[DownBox]) -> Dim:
    assert len(boxes) > 0
    if len(boxes) == 1:
        return boxes[0].dimensions()
    if len(boxes) <= 16:
        # mean
        width = 0
        height = 0
        for box in boxes:
            width += box.width()
            height += box.height()
        return (width // len(boxes), height // len(boxes))
    # median
    widths = [box.width() for box in boxes]
    heights = [box.height() for box in boxes]
    widths.sort()
    heights.sort()
    return (widths[len(widths) // 2], heights[len(heights) // 2])


def get_trimmed(boxes: List[DownBox]) -> Tuple[Box, Box]:
    avg = average_box_dimensions(boxes)

    outlier_dist = 1
    # threshold = 8
    # if avg[0] > threshold and avg[1] > threshold:
    #     outlier_dist = 2
    # threshold = 32
    # if avg[0] > threshold and avg[1] > threshold:
    #     outlier_dist = 3

    def is_outlier(box: DownBox) -> bool:
        dim = box.dimensions()
        if abs(dim[0] - avg[0]) > outlier_dist:
            return True
        if abs(dim[1] - avg[1]) > outlier_dist:
            return True
        return False
    
    assert len(boxes) > 0
    front = boxes[0]
    back = boxes[-1]

    min_out = (0, 0)
    max_out = back.max()
    min_down = (0, 0)
    max_down = back.down_pos()
    if is_outlier(front):
        for i in range(1, len(boxes)):
            if not is_outlier(boxes[i]):
                min_out = boxes[i].min()
                min_down = boxes[i].down_pos()
                break
    if is_outlier(back):
        for i in range(len(boxes) - 2, -1, -1):
            if not is_outlier(boxes[i]):
                max_out = boxes[i].max()
                max_down = boxes[i].down_pos()
                break
    box_out = Box(min_out, max_out)
    box_down = Box(min_down, max_down)
    return (box_out, box_down)


def calc_face_box(control_image: Image.Image, min_pos: Pos) -> Box:
    min_pixel = control_image.getpixel(min_pos)
    width, height = control_image.size
    x = 0
    while min_pos[0] + x < width:
        if control_image.getpixel((min_pos[0] + x, min_pos[1])) != min_pixel:
            break
        x += 1
    y = 0
    while min_pos[1] + y < height:
        if control_image.getpixel((min_pos[0], min_pos[1] + y)) != min_pixel:
            break
        y += 1
    x -= 1
    y -= 1
    assert x > 0
    assert y > 0
    return Box(min_pos, (x + min_pos[0], y + min_pos[1]))


def extract_boxes(control_image: Image.Image) -> ExtractedBoxes:
    width, height = control_image.size
    assert width > 0
    assert height > 0

    boxes: List[DownBox] = []
    x = 0
    y = 0
    down_x = 0
    down_y = 0

    while y < height:
        while x < width:
            min_pos = (x, y)
            box = calc_face_box(control_image, min_pos)
            boxes.append(DownBox(box.min(), box.max(), (down_x, down_y)))
            x += box.width()
            down_x += 1
        assert x == width
        box = boxes[-1]
        x = 0
        y += box.height()
        down_x = 0
        down_y += 1
    assert y == height

    return ExtractedBoxes(boxes)


def downsample_one(input_image: Image.Image, box: Box, sample_radius: Optional[int], downsampler: Image.Resampling) -> Tuple[int, int, int]:
    region = input_image.crop(box.as_tuple())

    box_width = box.width()
    box_height = box.height()
    box_center_x = box.min()[0] + box_width // 2
    box_center_y = box.min()[1] + box_height // 2

    if sample_radius is not None:
        radius_x = min(sample_radius, box_width // 2)
        radius_y = min(sample_radius, box_height // 2)
    else:
        radius_x = box_width // 2
        radius_y = box_height // 2

    cropped_region = region.crop((
        max(0, box_center_x - radius_x - box.min()[0]),
        max(0, box_center_y - radius_y - box.min()[1]),
        min(box_width, box_center_x + radius_x - box.min()[0]),
        min(box_height, box_center_y + radius_y - box.min()[1])
    ))
    assert cropped_region.size[0] >= radius_x and cropped_region.size[1] >= radius_y
    sampled = cropped_region.resize((1, 1), downsampler)

    rgb_value = sampled.getpixel((0, 0))
    assert isinstance(rgb_value, tuple) and len(rgb_value) == 3
    return rgb_value


class ImageRef:
    def __init__(self, ref: Image.Image) -> None:
        self.ref = ref


def downsample_all(*, input_image: Image.Image, output_image: Optional[ImageRef], down_image: Optional[ImageRef], boxes: List[DownBox], sample_radius: Optional[int], downsampler: Image.Resampling, trim_cropped_edges: bool) -> None:
    assert output_image or down_image
    for box in boxes:
        rgb_value = downsample_one(input_image, box, sample_radius, downsampler)
        solid_color_image = Image.new("RGB", box.dimensions(), rgb_value)
        if output_image:
            output_image.ref.paste(solid_color_image, box.min())
        if down_image:
            down_image.ref.paste(solid_color_image, box.down_pos())
    if trim_cropped_edges:
        o, d = get_trimmed(boxes)
        if output_image:
            output_image.ref = output_image.ref.crop(o.as_tuple())
        if down_image:
            down_image.ref = down_image.ref.crop(d.as_tuple())


def str2bool(value) -> bool:
    if isinstance(value, bool):
       return value
    if value.lower() in ("true", "1"):
        return True
    elif value.lower() in ("false", "0"):
        return False
    else:
        raise argparse.ArgumentTypeError("Boolean value expected.")


def controlled_downscale(*, control_path: str, input_path: str, output_downscaled_path: Optional[str], output_quantized_path: Optional[str], sample_radius: Optional[int], downsampler: Image.Resampling, trim_cropped_edges: bool, output_colorized_full_path: Optional[str], output_colorized_down_path: Optional[str]) -> None:
    """
    Downsample and rescale an image.
 
    :param control_path: Path to the control image.
    :param input_path: Path to the input image.
    :param output_downscaled_path: Path to save the output downscaled image.
    :param output_quantized_path: Path to save the output quantized image (downscaled and then upscaled to the original size).
    :param sample_radius: Radius for sampling (Manhattan distance).
    :param downsampler: Downsampler to use.
    :param trim_cropped_edges: Drop mapped checker grid elements that are cropped in the control image.
    :param output_colorized_full_path: Colorize the full checker image to debug the checker parsing.
    :param output_colorized_down_path: Colorize the downscaled checker image to debug the checker parsing.
    """
    if not output_downscaled_path and not output_quantized_path:
        raise ValueError("At least one of output_up and output_down must be specified.")
 
    control_image = Image.open(control_path).convert("1")
    input_image = Image.open(input_path)
    if control_image.size != input_image.size:
        raise ValueError("Control image and input image must have the same dimensions.")

    downscaled_image: Optional[ImageRef] = None
    quantized_image: Optional[ImageRef] = None
 
    if output_quantized_path:
        quantized_image = ImageRef(Image.new("RGB", input_image.size))

    extracted_boxes = extract_boxes(control_image)
    if output_colorized_full_path:
        extracted_boxes.to_colored_checkers(full=True).save(output_colorized_full_path)
    if output_colorized_down_path:
        extracted_boxes.to_colored_checkers(full=False).save(output_colorized_down_path)
 
    if output_downscaled_path:
        downscaled_image = ImageRef(Image.new("RGB", extracted_boxes.down_dimensions()))

    boxes = extracted_boxes.boxes()
    downsample_all(input_image=input_image, output_image=quantized_image, down_image=downscaled_image, boxes=boxes, sample_radius=sample_radius, downsampler=downsampler, trim_cropped_edges=trim_cropped_edges)
 
    if quantized_image:
        assert output_quantized_path
        quantized_image.ref.save(output_quantized_path)
    if downscaled_image:
        assert output_downscaled_path
        downscaled_image.ref.save(output_downscaled_path)


def main(cli_args: List[str]) -> None:
    parser = argparse.ArgumentParser(description="Downsample and rescale image.")
    parser.add_argument("--control", type=str, required=True, help="Path to control image.")
    parser.add_argument("--input", type=str, required=True, help="Path to input image.")
    parser.add_argument("--output-downscaled", type=str, help="Path to save the output downscaled image.")
    parser.add_argument("--output-quantized", type=str, help="Path to save the output quantized image (downscaled and then upscaled to the original size).")
    parser.add_argument("--sample-radius", type=int, default=None, help="Radius for sampling (Manhattan distance).")
    parser.add_argument("--downsampler", choices=["box", "bilinear", "bicubic", "hamming", "lanczos"], default="box", help="Downsampler to use.")
    parser.add_argument("--trim-cropped-edges", type=str2bool, default=False, help="Drop mapped checker grid elements that are cropped in the control image.")
    parser.add_argument("--output-colorized-full", type=str, help="Colorize the full checker image to debug the checker parsing.")
    parser.add_argument("--output-colorized-down", type=str, help="Colorize the downscaled checker image to debug the checker parsing.")

    args = parser.parse_args(cli_args)
    downsampler = Image.Resampling[args.downsampler.upper()]
    
    controlled_downscale(
        control_path=args.control,
        input_path=args.input,
        output_downscaled_path=args.output_downscaled,
        output_quantized_path=args.output_quantized,
        sample_radius=args.sample_radius,
        downsampler=downsampler,
        trim_cropped_edges=args.trim_cropped_edges,
        output_colorized_full_path=args.output_colorized_full,
        output_colorized_down_path=args.output_colorized_down,
    )


if __name__ == "__main__":
    main(sys.argv[1:])