Functional transforms (augmentations.functional)¶

@uint8_io
@clipped
@preserve_channel_dim
def add_fog(
    img: np.ndarray,
    fog_intensity: float,
    alpha_coef: float,
    fog_particle_positions: list[tuple[int, int]],
    random_state: np.random.RandomState | None = None,
) -> np.ndarray:
    """Add fog to the input image.

    Args:
        img (np.ndarray): Input image.
        fog_intensity (float): Intensity of the fog effect, between 0 and 1.
        alpha_coef (float): Base alpha (transparency) value for fog particles.
        fog_particle_positions (list[tuple[int, int]]): List of (x, y) coordinates for fog particles.
        random_state (np.random.RandomState | None): If specified, this will be random state used
    Returns:
        np.ndarray: Image with added fog effect.
    """
    height, width = img.shape[:2]
    num_channels = get_num_channels(img)

    fog_layer = np.zeros((height, width, num_channels), dtype=np.uint8)

    max_fog_radius = int(
        min(height, width) * 0.1 * fog_intensity,
    )  # Maximum radius scales with image size and intensity

    for x, y in fog_particle_positions:
        radius = random_utils.randint(max_fog_radius // 2, max_fog_radius, random_state=random_state)
        color = 255 if num_channels == 1 else (255,) * num_channels
        cv2.circle(
            fog_layer,
            center=(x, y),
            radius=radius,
            color=color,
            thickness=-1,
        )

    # Apply gaussian blur to the fog layer
    fog_layer = cv2.GaussianBlur(fog_layer, (25, 25), 0)

    # Blend the fog layer with the original image
    alpha = np.mean(fog_layer, axis=2, keepdims=True) / 255 * alpha_coef * fog_intensity
    fog_image = img * (1 - alpha) + fog_layer * alpha

    return fog_image.astype(np.uint8)

@uint8_io
@preserve_channel_dim
def add_rain(
    img: np.ndarray,
    slant: int,
    drop_length: int,
    drop_width: int,
    drop_color: tuple[int, int, int],
    blur_value: int,
    brightness_coefficient: float,
    rain_drops: list[tuple[int, int]],
) -> np.ndarray:
    """Adds rain drops to the image.

    Args:
        img (np.ndarray): Input image.
        slant (int): The angle of the rain drops.
        drop_length (int): The length of each rain drop.
        drop_width (int): The width of each rain drop.
        drop_color (tuple[int, int, int]): The color of the rain drops in RGB format.
        blur_value (int): The size of the kernel used to blur the image. Rainy views are blurry.
        brightness_coefficient (float): Coefficient to adjust the brightness of the image. Rainy days are usually shady.
        rain_drops (list[tuple[int, int]]): A list of tuples where each tuple represents the (x, y)
            coordinates of the starting point of a rain drop.

    Returns:
        np.ndarray: Image with rain effect added.

    Reference:
        https://github.com/UjjwalSaxena/Automold--Road-Augmentation-Library
    """
    for rain_drop_x0, rain_drop_y0 in rain_drops:
        rain_drop_x1 = rain_drop_x0 + slant
        rain_drop_y1 = rain_drop_y0 + drop_length

        cv2.line(
            img,
            (rain_drop_x0, rain_drop_y0),
            (rain_drop_x1, rain_drop_y1),
            drop_color,
            drop_width,
        )

    img = cv2.blur(img, (blur_value, blur_value))  # rainy view are blurry
    image_hsv = cv2.cvtColor(img, cv2.COLOR_RGB2HSV).astype(np.float32)
    image_hsv[:, :, 2] *= brightness_coefficient

    return cv2.cvtColor(image_hsv.astype(np.uint8), cv2.COLOR_HSV2RGB)

@uint8_io
@preserve_channel_dim
def add_shadow(img: np.ndarray, vertices_list: list[np.ndarray], intensities: np.ndarray) -> np.ndarray:
    """Add shadows to the image by reducing the intensity of the pixel values in specified regions.

    Args:
        img (np.ndarray): Input image. Multichannel images are supported.
        vertices_list (list[np.ndarray]): List of vertices for shadow polygons.
        intensities (np.ndarray): Array of shadow intensities. Range is [0, 1].

    Returns:
        np.ndarray: Image with shadows added.

    Reference:
        https://github.com/UjjwalSaxena/Automold--Road-Augmentation-Library
    """
    num_channels = get_num_channels(img)
    max_value = MAX_VALUES_BY_DTYPE[np.uint8]

    img_shadowed = img.copy()

    # Iterate over the vertices and intensity list
    for vertices, shadow_intensity in zip(vertices_list, intensities):
        # Create mask for the current shadow polygon
        mask = np.zeros((img.shape[0], img.shape[1], 1), dtype=np.uint8)
        cv2.fillPoly(mask, [vertices], (max_value,))

        # Duplicate the mask to have the same number of channels as the image
        mask = np.repeat(mask, num_channels, axis=2)

        # Apply shadow to the channels directly
        # It could be tempting to convert to HLS and apply the shadow to the L channel, but it creates artifacts
        shadowed_indices = mask[:, :, 0] == max_value
        img_shadowed[shadowed_indices] = clip(
            img_shadowed[shadowed_indices] * shadow_intensity,
            np.uint8,
        )

    return img_shadowed

@uint8_io
def add_snow_bleach(img: np.ndarray, snow_point: float, brightness_coeff: float) -> np.ndarray:
    """Adds a simple snow effect to the image by bleaching out pixels.

    This function simulates a basic snow effect by increasing the brightness of pixels
    that are above a certain threshold (snow_point). It operates in the HLS color space
    to modify the lightness channel.

    Args:
        img (np.ndarray): Input image. Can be either RGB uint8 or float32.
        snow_point (float): A float in the range [0, 1], scaled and adjusted to determine
            the threshold for pixel modification. Higher values result in less snow effect.
        brightness_coeff (float): Coefficient applied to increase the brightness of pixels
            below the snow_point threshold. Larger values lead to more pronounced snow effects.
            Should be greater than 1.0 for a visible effect.

    Returns:
        np.ndarray: Image with simulated snow effect. The output has the same dtype as the input.

    Note:
        - This function converts the image to the HLS color space to modify the lightness channel.
        - The snow effect is created by selectively increasing the brightness of pixels.
        - This method tends to create a 'bleached' look, which may not be as realistic as more
          advanced snow simulation techniques.
        - The function automatically handles both uint8 and float32 input images.

    The snow effect is created through the following steps:
    1. Convert the image from RGB to HLS color space.
    2. Adjust the snow_point threshold.
    3. Increase the lightness of pixels below the threshold.
    4. Convert the image back to RGB.

    Mathematical Formulation:
        Let L be the lightness channel in HLS space.
        For each pixel (i, j):
        If L[i, j] < snow_point:
            L[i, j] = L[i, j] * brightness_coeff

    Examples:
        >>> import numpy as np
        >>> import albumentations as A
        >>> image = np.random.randint(0, 256, [100, 100, 3], dtype=np.uint8)
        >>> snowy_image = A.functional.add_snow_v1(image, snow_point=0.5, brightness_coeff=1.5)

    References:
        - HLS Color Space: https://en.wikipedia.org/wiki/HSL_and_HSV
        - Original implementation: https://github.com/UjjwalSaxena/Automold--Road-Augmentation-Library
    """
    max_value = MAX_VALUES_BY_DTYPE[np.uint8]

    snow_point *= max_value / 2
    snow_point += max_value / 3

    image_hls = cv2.cvtColor(img, cv2.COLOR_RGB2HLS)
    image_hls = np.array(image_hls, dtype=np.float32)

    image_hls[:, :, 1][image_hls[:, :, 1] < snow_point] *= brightness_coeff

    image_hls[:, :, 1] = clip(image_hls[:, :, 1], np.uint8)

    image_hls = np.array(image_hls, dtype=np.uint8)

    return cv2.cvtColor(image_hls, cv2.COLOR_HLS2RGB)

@uint8_io
def add_snow_texture(img: np.ndarray, snow_point: float, brightness_coeff: float) -> np.ndarray:
    """Add a realistic snow effect to the input image.

    This function simulates snowfall by applying multiple visual effects to the image,
    including brightness adjustment, snow texture overlay, depth simulation, and color tinting.
    The result is a more natural-looking snow effect compared to simple pixel bleaching methods.

    Args:
        img (np.ndarray): Input image in RGB format.
        snow_point (float): Coefficient that controls the amount and intensity of snow.
            Should be in the range [0, 1], where 0 means no snow and 1 means maximum snow effect.
        brightness_coeff (float): Coefficient for brightness adjustment to simulate the
            reflective nature of snow. Should be in the range [0, 1], where higher values
            result in a brighter image.

    Returns:
        np.ndarray: Image with added snow effect. The output has the same dtype as the input.

    Note:
        - The function first converts the image to HSV color space for better control over
          brightness and color adjustments.
        - A snow texture is generated using Gaussian noise and then filtered for a more
          natural appearance.
        - A depth effect is simulated, with more snow at the top of the image and less at the bottom.
        - A slight blue tint is added to simulate the cool color of snow.
        - Random sparkle effects are added to simulate light reflecting off snow crystals.

    The snow effect is created through the following steps:
    1. Brightness adjustment in HSV space
    2. Generation of a snow texture using Gaussian noise
    3. Application of a depth effect to the snow texture
    4. Blending of the snow texture with the original image
    5. Addition of a cool blue tint
    6. Addition of sparkle effects

    Examples:
        >>> import numpy as np
        >>> import albumentations as A
        >>> image = np.random.randint(0, 256, [100, 100, 3], dtype=np.uint8)
        >>> snowy_image = A.functional.add_snow_v2(image, snow_coeff=0.5, brightness_coeff=0.2)

    Note:
        This function works with both uint8 and float32 image types, automatically
        handling the conversion between them.

    References:
        - Perlin Noise: https://en.wikipedia.org/wiki/Perlin_noise
        - HSV Color Space: https://en.wikipedia.org/wiki/HSL_and_HSV
    """
    max_value = MAX_VALUES_BY_DTYPE[np.uint8]

    # Convert to HSV for better color control
    img_hsv = cv2.cvtColor(img, cv2.COLOR_RGB2HSV).astype(np.float32)

    # Increase brightness
    img_hsv[:, :, 2] = np.clip(img_hsv[:, :, 2] * (1 + brightness_coeff * snow_point), 0, max_value)

    # Generate snow texture
    snow_texture = random_utils.normal(size=img.shape[:2], loc=0.5, scale=0.3)
    snow_texture = cv2.GaussianBlur(snow_texture, (0, 0), sigmaX=1, sigmaY=1)

    # Create depth effect for snow simulation
    # More snow accumulates at the top of the image, gradually decreasing towards the bottom
    # This simulates natural snow distribution on surfaces
    # The effect is achieved using a linear gradient from 1 (full snow) to 0.2 (less snow)
    rows = img.shape[0]
    depth_effect = np.linspace(1, 0.2, rows)[:, np.newaxis]
    snow_texture *= depth_effect

    # Apply snow texture
    snow_layer = (np.dstack([snow_texture] * 3) * max_value * snow_point).astype(np.float32)

    # Blend snow with original image
    img_with_snow = cv2.addWeighted(img_hsv, 1, snow_layer, 1, 0)

    # Add a slight blue tint to simulate cool snow color
    blue_tint = np.full_like(img_with_snow, (0.6, 0.75, 1))  # Slight blue in HSV

    img_with_snow = cv2.addWeighted(img_with_snow, 0.85, blue_tint, 0.15 * snow_point, 0)

    # Convert back to RGB
    img_with_snow = cv2.cvtColor(img_with_snow.astype(np.uint8), cv2.COLOR_HSV2RGB)

    # Add some sparkle effects for snow glitter
    sparkle = random_utils.random(img.shape[:2]) > 0.99  # noqa: PLR2004
    img_with_snow[sparkle] = [max_value, max_value, max_value]

    return img_with_snow

@uint8_io
@preserve_channel_dim
def add_sun_flare_overlay(
    img: np.ndarray,
    flare_center: tuple[float, float],
    src_radius: int,
    src_color: ColorType,
    circles: list[Any],
) -> np.ndarray:
    """Add a sun flare effect to an image using a simple overlay technique.

    This function creates a basic sun flare effect by overlaying multiple semi-transparent
    circles of varying sizes and intensities on the input image. The effect simulates
    a simple lens flare caused by bright light sources.

    Args:
        img (np.ndarray): The input image.
        flare_center (tuple[float, float]): (x, y) coordinates of the flare center
            in pixel coordinates.
        src_radius (int): The radius of the main sun circle in pixels.
        src_color (ColorType): The color of the sun, represented as a tuple of RGB values.
        circles (list[Any]): A list of tuples, each representing a circle that contributes
            to the flare effect. Each tuple contains:
            - alpha (float): The transparency of the circle (0.0 to 1.0).
            - center (tuple[int, int]): (x, y) coordinates of the circle center.
            - radius (int): The radius of the circle.
            - color (tuple[int, int, int]): RGB color of the circle.

    Returns:
        np.ndarray: The output image with the sun flare effect added.

    Note:
        - This function uses a simple alpha blending technique to overlay flare elements.
        - The main sun is created as a gradient circle, fading from the center outwards.
        - Additional flare circles are added along an imaginary line from the sun's position.
        - This method is computationally efficient but may produce less realistic results
          compared to more advanced techniques.

    The flare effect is created through the following steps:
    1. Create an overlay image and output image as copies of the input.
    2. Add smaller flare circles to the overlay.
    3. Blend the overlay with the output image using alpha compositing.
    4. Add the main sun circle with a radial gradient.

    Examples:
        >>> import numpy as np
        >>> import albumentations as A
        >>> image = np.random.randint(0, 256, [100, 100, 3], dtype=np.uint8)
        >>> flare_center = (50, 50)
        >>> src_radius = 20
        >>> src_color = (255, 255, 200)
        >>> circles = [
        ...     (0.1, (60, 60), 5, (255, 200, 200)),
        ...     (0.2, (70, 70), 3, (200, 255, 200))
        ... ]
        >>> flared_image = A.functional.add_sun_flare_overlay(
        ...     image, flare_center, src_radius, src_color, circles
        ... )

    References:
        - Alpha compositing: https://en.wikipedia.org/wiki/Alpha_compositing
        - Lens flare: https://en.wikipedia.org/wiki/Lens_flare
    """
    overlay = img.copy()
    output = img.copy()

    for alpha, (x, y), rad3, (r_color, g_color, b_color) in circles:
        cv2.circle(overlay, (x, y), rad3, (r_color, g_color, b_color), -1)
        output = add_weighted(overlay, alpha, output, 1 - alpha)

    point = [int(x) for x in flare_center]

    overlay = output.copy()
    num_times = src_radius // 10
    alpha = np.linspace(0.0, 1, num=num_times)
    rad = np.linspace(1, src_radius, num=num_times)
    for i in range(num_times):
        cv2.circle(overlay, point, int(rad[i]), src_color, -1)
        alp = alpha[num_times - i - 1] * alpha[num_times - i - 1] * alpha[num_times - i - 1]
        output = add_weighted(overlay, alp, output, 1 - alp)

    return output

@uint8_io
@clipped
def add_sun_flare_physics_based(
    img: np.ndarray,
    flare_center: tuple[int, int],
    src_radius: int,
    src_color: tuple[int, int, int],
    circles: list[Any],
) -> np.ndarray:
    """Add a more realistic sun flare effect to the image.

    This function creates a complex sun flare effect by simulating various optical phenomena
    that occur in real camera lenses when capturing bright light sources. The result is a
    more realistic and physically plausible lens flare effect.

    Args:
        img (np.ndarray): Input image.
        flare_center (tuple[int, int]): (x, y) coordinates of the sun's center in pixels.
        src_radius (int): Radius of the main sun circle in pixels.
        src_color (tuple[int, int, int]): Color of the sun in RGB format.
        circles (list[Any]): List of tuples, each representing a flare circle with parameters:
            (alpha, center, size, color)
            - alpha (float): Transparency of the circle (0.0 to 1.0).
            - center (tuple[int, int]): (x, y) coordinates of the circle center.
            - size (float): Size factor for the circle radius.
            - color (tuple[int, int, int]): RGB color of the circle.

    Returns:
        np.ndarray: Image with added sun flare effect.

    Note:
        This function implements several techniques to create a more realistic flare:
        1. Separate flare layer: Allows for complex manipulations of the flare effect.
        2. Lens diffraction spikes: Simulates light diffraction in camera aperture.
        3. Radial gradient mask: Creates natural fading of the flare from the center.
        4. Gaussian blur: Softens the flare for a more natural glow effect.
        5. Chromatic aberration: Simulates color fringing often seen in real lens flares.
        6. Screen blending: Provides a more realistic blending of the flare with the image.

    The flare effect is created through the following steps:
    1. Create a separate flare layer.
    2. Add the main sun circle and diffraction spikes to the flare layer.
    3. Add additional flare circles based on the input parameters.
    4. Apply Gaussian blur to soften the flare.
    5. Create and apply a radial gradient mask for natural fading.
    6. Simulate chromatic aberration by applying different blurs to color channels.
    7. Blend the flare with the original image using screen blending mode.

    Examples:
        >>> import numpy as np
        >>> import albumentations as A
        >>> image = np.random.randint(0, 256, [1000, 1000, 3], dtype=np.uint8)
        >>> flare_center = (500, 500)
        >>> src_radius = 50
        >>> src_color = (255, 255, 200)
        >>> circles = [
        ...     (0.1, (550, 550), 10, (255, 200, 200)),
        ...     (0.2, (600, 600), 5, (200, 255, 200))
        ... ]
        >>> flared_image = A.functional.add_sun_flare_physics_based(
        ...     image, flare_center, src_radius, src_color, circles
        ... )

    References:
        - Lens flare: https://en.wikipedia.org/wiki/Lens_flare
        - Diffraction: https://en.wikipedia.org/wiki/Diffraction
        - Chromatic aberration: https://en.wikipedia.org/wiki/Chromatic_aberration
        - Screen blending: https://en.wikipedia.org/wiki/Blend_modes#Screen
    """
    output = img.copy()
    height, width = img.shape[:2]

    # Create a separate flare layer
    flare_layer = np.zeros_like(img, dtype=np.float32)

    # Add the main sun
    cv2.circle(flare_layer, flare_center, src_radius, src_color, -1)

    # Add lens diffraction spikes
    for angle in [0, 45, 90, 135]:
        end_point = (
            int(flare_center[0] + np.cos(np.radians(angle)) * max(width, height)),
            int(flare_center[1] + np.sin(np.radians(angle)) * max(width, height)),
        )
        cv2.line(flare_layer, flare_center, end_point, src_color, 2)

    # Add flare circles
    for _, center, size, color in circles:
        cv2.circle(flare_layer, center, int(size**0.33), color, -1)

    # Apply gaussian blur to soften the flare
    flare_layer = cv2.GaussianBlur(flare_layer, (0, 0), sigmaX=15, sigmaY=15)

    # Create a radial gradient mask
    y, x = np.ogrid[:height, :width]
    mask = np.sqrt((x - flare_center[0]) ** 2 + (y - flare_center[1]) ** 2)
    mask = 1 - np.clip(mask / (max(width, height) * 0.7), 0, 1)
    mask = np.dstack([mask] * 3)

    # Apply the mask to the flare layer
    flare_layer *= mask

    # Add chromatic aberration
    channels = list(cv2.split(flare_layer))
    channels[0] = cv2.GaussianBlur(channels[0], (0, 0), sigmaX=3, sigmaY=3)  # Blue channel
    channels[2] = cv2.GaussianBlur(channels[2], (0, 0), sigmaX=5, sigmaY=5)  # Red channel
    flare_layer = cv2.merge(channels)

    # Blend the flare with the original image using screen blending
    return 255 - ((255 - output) * (255 - flare_layer) / 255)

def almost_equal_intervals(n: int, parts: int) -> np.ndarray:
    """Generates an array of nearly equal integer intervals that sum up to `n`.

    This function divides the number `n` into `parts` nearly equal parts. It ensures that
    the sum of all parts equals `n`, and the difference between any two parts is at most one.
    This is useful for distributing a total amount into nearly equal discrete parts.

    Args:
        n (int): The total value to be split.
        parts (int): The number of parts to split into.

    Returns:
        np.ndarray: An array of integers where each integer represents the size of a part.

    Example:
        >>> almost_equal_intervals(20, 3)
        array([7, 7, 6])  # Splits 20 into three parts: 7, 7, and 6
        >>> almost_equal_intervals(16, 4)
        array([4, 4, 4, 4])  # Splits 16 into four equal parts
    """
    part_size, remainder = divmod(n, parts)
    # Create an array with the base part size and adjust the first `remainder` parts by adding 1
    return np.array([part_size + 1 if i < remainder else part_size for i in range(parts)])

@uint8_io
@preserve_channel_dim
def clahe(img: np.ndarray, clip_limit: float, tile_grid_size: tuple[int, int]) -> np.ndarray:
    """Apply Contrast Limited Adaptive Histogram Equalization (CLAHE) to the input image.

    This function enhances the contrast of the input image using CLAHE. For color images,
    it converts the image to the LAB color space, applies CLAHE to the L channel, and then
    converts the image back to RGB.

    Args:
        img (np.ndarray): Input image. Can be grayscale (2D array) or RGB (3D array).
        clip_limit (float): Threshold for contrast limiting. Higher values give more contrast.
        tile_grid_size (tuple[int, int]): Size of grid for histogram equalization.
            Width and height of the grid.

    Returns:
        np.ndarray: Image with CLAHE applied. The output has the same dtype as the input.

    Note:
        - If the input image is float32, it's temporarily converted to uint8 for processing
          and then converted back to float32.
        - For color images, CLAHE is applied only to the luminance channel in the LAB color space.

    Raises:
        ValueError: If the input image is not 2D or 3D.

    Example:
        >>> import numpy as np
        >>> img = np.random.randint(0, 256, (100, 100, 3), dtype=np.uint8)
        >>> result = clahe(img, clip_limit=2.0, tile_grid_size=(8, 8))
        >>> assert result.shape == img.shape
        >>> assert result.dtype == img.dtype
    """
    img = img.copy()
    clahe_mat = cv2.createCLAHE(clipLimit=clip_limit, tileGridSize=tile_grid_size)

    if is_grayscale_image(img):
        return clahe_mat.apply(img)

    img = cv2.cvtColor(img, cv2.COLOR_RGB2LAB)

    img[:, :, 0] = clahe_mat.apply(img[:, :, 0])

    return cv2.cvtColor(img, cv2.COLOR_LAB2RGB)

def create_shape_groups(tiles: np.ndarray) -> dict[tuple[int, int], list[int]]:
    """Groups tiles by their shape and stores the indices for each shape."""
    shape_groups = defaultdict(list)
    for index, (start_y, start_x, end_y, end_x) in enumerate(tiles):
        shape = (end_y - start_y, end_x - start_x)
        shape_groups[shape].append(index)
    return shape_groups

@uint8_io
@preserve_channel_dim
def equalize(
    img: np.ndarray,
    mask: np.ndarray | None = None,
    mode: ImageMode = "cv",
    by_channels: bool = True,
) -> np.ndarray:
    """Apply histogram equalization to the input image.

    This function enhances the contrast of the input image by equalizing its histogram.
    It supports both grayscale and color images, and can operate on individual channels
    or on the luminance channel of the image.

    Args:
        img (np.ndarray): Input image. Can be grayscale (2D array) or RGB (3D array).
        mask (np.ndarray | None): Optional mask to apply the equalization selectively.
            If provided, must have the same shape as the input image. Default: None.
        mode (ImageMode): The backend to use for equalization. Can be either "cv" for
            OpenCV or "pil" for Pillow-style equalization. Default: "cv".
        by_channels (bool): If True, applies equalization to each channel independently.
            If False, converts the image to YCrCb color space and equalizes only the
            luminance channel. Only applicable to color images. Default: True.

    Returns:
        np.ndarray: Equalized image. The output has the same dtype as the input.

    Raises:
        ValueError: If the input image or mask have invalid shapes or types.

    Note:
        - If the input image is not uint8, it will be temporarily converted to uint8
          for processing and then converted back to its original dtype.
        - For color images, when by_channels=False, the image is converted to YCrCb
          color space, equalized on the Y channel, and then converted back to RGB.
        - The function preserves the original number of channels in the image.

    Example:
        >>> import numpy as np
        >>> import albumentations as A
        >>> image = np.random.randint(0, 256, (100, 100, 3), dtype=np.uint8)
        >>> equalized = A.equalize(image, mode="cv", by_channels=True)
        >>> assert equalized.shape == image.shape
        >>> assert equalized.dtype == image.dtype
    """
    _check_preconditions(img, mask, by_channels)

    function = _equalize_pil if mode == "pil" else _equalize_cv

    if is_grayscale_image(img):
        return function(img, _handle_mask(mask))

    if not by_channels:
        result_img = cv2.cvtColor(img, cv2.COLOR_RGB2YCrCb)
        result_img[..., 0] = function(result_img[..., 0], _handle_mask(mask))
        return cv2.cvtColor(result_img, cv2.COLOR_YCrCb2RGB)

    result_img = np.empty_like(img)
    for i in range(NUM_RGB_CHANNELS):
        _mask = _handle_mask(mask, i)
        result_img[..., i] = function(img[..., i], _mask)

    return result_img

@float32_io
@clipped
@preserve_channel_dim
def fancy_pca(img: np.ndarray, alpha_vector: np.ndarray) -> np.ndarray:
    """Perform 'Fancy PCA' augmentation on an image with any number of channels.

    Args:
        img (np.ndarray): Input image
        alpha_vector (np.ndarray): Vector of scale factors for each principal component.
                                   Should have the same length as the number of channels in the image.

    Returns:
        np.ndarray: Augmented image of the same shape, type, and range as the input.

    Image types:
        uint8, float32

    Number of channels:
        Any

    Note:
        - This function generalizes the Fancy PCA augmentation to work with any number of channels.
        - It preserves the original range of the image ([0, 255] for uint8, [0, 1] for float32).
        - For single-channel images, the augmentation is applied as a simple scaling of pixel intensity variation.
        - For multi-channel images, PCA is performed on the entire image, treating each pixel
          as a point in N-dimensional space (where N is the number of channels).
        - The augmentation preserves the correlation between channels while adding controlled noise.
        - Computation time may increase significantly for images with a large number of channels.

    Reference:
        Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012).
        ImageNet classification with deep convolutional neural networks.
        In Advances in neural information processing systems (pp. 1097-1105).
    """
    orig_shape = img.shape
    num_channels = get_num_channels(img)

    # Reshape image to 2D array of pixels
    img_reshaped = img.reshape(-1, num_channels)

    # Center the pixel values
    img_mean = np.mean(img_reshaped, axis=0)
    img_centered = img_reshaped - img_mean

    if num_channels == 1:
        # For grayscale images, apply a simple scaling
        std_dev = np.std(img_centered)
        noise = alpha_vector[0] * std_dev * img_centered
    else:
        # Compute covariance matrix
        img_cov = np.cov(img_centered, rowvar=False)

        # Compute eigenvectors & eigenvalues of the covariance matrix
        eig_vals, eig_vecs = np.linalg.eigh(img_cov)

        # Sort eigenvectors by eigenvalues in descending order
        sort_perm = eig_vals[::-1].argsort()
        eig_vals = eig_vals[sort_perm]
        eig_vecs = eig_vecs[:, sort_perm]

        # Create noise vector
        noise = np.dot(np.dot(eig_vecs, np.diag(alpha_vector * eig_vals)), img_centered.T).T

    # Add noise to the image
    img_pca = img_reshaped + noise

    # Reshape back to original shape
    img_pca = img_pca.reshape(orig_shape)

    # Clip values to [0, 1] range
    return np.clip(img_pca, 0, 1)

def generate_shuffled_splits(
    size: int,
    divisions: int,
    random_state: np.random.RandomState | None = None,
) -> np.ndarray:
    """Generate shuffled splits for a given dimension size and number of divisions.

    Args:
        size (int): Total size of the dimension (height or width).
        divisions (int): Number of divisions (rows or columns).
        random_state (Optional[np.random.RandomState]): Seed for the random number generator for reproducibility.

    Returns:
        np.ndarray: Cumulative edges of the shuffled intervals.
    """
    intervals = almost_equal_intervals(size, divisions)
    intervals = random_utils.shuffle(intervals, random_state=random_state)
    return np.insert(np.cumsum(intervals), 0, 0)

def grayscale_to_multichannel(grayscale_image: np.ndarray, num_output_channels: int = 3) -> np.ndarray:
    """Convert a grayscale image to a multi-channel image.

    This function takes a 2D grayscale image or a 3D image with a single channel
    and converts it to a multi-channel image by repeating the grayscale data
    across the specified number of channels.

    Args:
        grayscale_image (np.ndarray): Input grayscale image. Can be 2D (height, width)
                                      or 3D (height, width, 1).
        num_output_channels (int, optional): Number of channels in the output image. Defaults to 3.

    Returns:
        np.ndarray: Multi-channel image with shape (height, width, num_channels).

    Note:
        If the input is already a multi-channel image with the desired number of channels,
        it will be returned unchanged.
    """
    grayscale_image = grayscale_image.copy().squeeze()
    return np.stack([grayscale_image] * num_output_channels, axis=-1)

@float32_io
@clipped
def iso_noise(
    image: np.ndarray,
    color_shift: float = 0.05,
    intensity: float = 0.5,
    random_state: np.random.RandomState | None = None,
) -> np.ndarray:
    """Apply poisson noise to an image to simulate camera sensor noise.

    Args:
        image (np.ndarray): Input image. Currently, only RGB images are supported.
        color_shift (float): The amount of color shift to apply. Default is 0.05.
        intensity (float): Multiplication factor for noise values. Values of ~0.5 produce a noticeable,
                           yet acceptable level of noise. Default is 0.5.
        random_state (np.random.RandomState | None): If specified, this will be random state used
            for noise generation.

    Returns:
        np.ndarray: The noised image.

    Image types:
        uint8, float32

    Number of channels:
        3
    """
    hls = cv2.cvtColor(image, cv2.COLOR_RGB2HLS)
    _, stddev = cv2.meanStdDev(hls)

    luminance_noise = random_utils.poisson(stddev[1] * intensity * 255, size=hls.shape[:2], random_state=random_state)
    color_noise = random_utils.normal(0, color_shift * 360 * intensity, size=hls.shape[:2], random_state=random_state)

    hue = hls[..., 0]
    hue += color_noise
    hue %= 360

    luminance = hls[..., 1]
    luminance += (luminance_noise / 255) * (1.0 - luminance)

    return cv2.cvtColor(hls, cv2.COLOR_HLS2RGB)

@uint8_io
@preserve_channel_dim
def move_tone_curve(
    img: np.ndarray,
    low_y: float | np.ndarray,
    high_y: float | np.ndarray,
) -> np.ndarray:
    """Rescales the relationship between bright and dark areas of the image by manipulating its tone curve.

    Args:
        img: np.ndarray. Any number of channels
        low_y: per-channel or single y-position of a Bezier control point used
            to adjust the tone curve, must be in range [0, 1]
        high_y: per-channel or single y-position of a Bezier control point used
            to adjust image tone curve, must be in range [0, 1]

    """
    t = np.linspace(0.0, 1.0, 256)

    def evaluate_bez(t: np.ndarray, low_y: float | np.ndarray, high_y: float | np.ndarray) -> np.ndarray:
        one_minus_t = 1 - t
        return (3 * one_minus_t**2 * t * low_y + 3 * one_minus_t * t**2 * high_y + t**3) * 255

    num_channels = get_num_channels(img)

    if np.isscalar(low_y) and np.isscalar(high_y):
        lut = clip(np.rint(evaluate_bez(t, low_y, high_y)), np.uint8)
        return cv2.LUT(img, lut)
    if isinstance(low_y, np.ndarray) and isinstance(high_y, np.ndarray):
        luts = clip(np.rint(evaluate_bez(t[:, np.newaxis], low_y, high_y).T), np.uint8)
        return cv2.merge([cv2.LUT(img[:, :, i], luts[i]) for i in range(num_channels)])

    raise TypeError(
        f"low_y and high_y must both be of type float or np.ndarray. Got {type(low_y)} and {type(high_y)}",
    )