Colorize

Targets:

image

Image Types:uint8, float32

Map a single-channel grayscale image to a 2- or 3-color RGB gradient with per-call sampled anchor colors (Pillow ImageOps.colorize style).

Intensity acts as a coordinate along a sampled color ramp:

0 maps to a sample from black_range
255 (or 1.0 for float32) maps to a sample from white_range
if mid_range is set, intensity sampled from mid_value_range maps to a sample from mid_range and the ramp becomes piecewise linear

Each anchor range is given as (low_rgb, high_rgb) and sampled per-channel uniformly on every call. Pass identical low/high tuples to fix a color (e.g. black_range=((0, 0, 255), (0, 0, 255))). Anchors are always specified in 0-255 RGB; for float32 inputs they are rescaled to [0, 1] internally.

Arguments

black_range

tuple[tuple[int, int, int], tuple[int, int, int]]

[[0,0,0],[0,0,0]]

Inclusive per-channel range from which the dark anchor is sampled. Default: ((0, 0, 0), (0, 0, 0)).

white_range

tuple[tuple[int, int, int], tuple[int, int, int]]

[[255,255,255],[255,255,255]]

Inclusive per-channel range from which the bright anchor is sampled. Default: ((255, 255, 255), (255, 255, 255)).

mid_range

tuple[tuple[int, int, int], tuple[int, int, int]] | None

Optional inclusive range from which the midpoint anchor is sampled. None disables 3-color mode. Default: None.

mid_value_range

tuple[int, int]

[127,127]

Inclusive intensity range (each in 1-254) from which the midpoint position is sampled. Ignored when mid is None. Default: (127, 127).

p

float

0.5

Probability of applying the transform. Default: 0.5.

Examples

>>> import numpy as np
>>> import albumentations as A
>>> image = np.random.randint(0, 256, (100, 100, 1), dtype=np.uint8)
>>>
>>> # Fixed blue -> yellow ramp (low == high)
>>> fixed = A.Compose([A.Colorize(
...     black_range=((0, 0, 255), (0, 0, 255)),
...     white_range=((255, 255, 0), (255, 255, 0)),
...     p=1.0,
... )])
>>> assert fixed(image=image)["image"].shape == (100, 100, 3)
>>>
>>> # Random thermal-ish ramp with random midpoint position
>>> random_thermal = A.Compose([A.Colorize(
...     black_range=((0, 0, 64), (32, 0, 192)),
...     mid_range=((96, 0, 96), (160, 64, 160)),
...     white_range=((220, 160, 0), (255, 220, 32)),
...     mid_value_range=(96, 160),
...     p=1.0,
... )])
>>> assert random_thermal(image=image)["image"].shape == (100, 100, 3)

Notes

Input must be single-channel; multi-channel input is a no-op with a warning.
Interpolation is linear in RGB space.
For uint8 inputs the per-call mapping is a (256, 3) LUT applied via cv2.LUT; for float32 inputs np.interp is used per channel.

>>> import numpy as np >>> import albumentations as A >>> image = np.random.randint(0, 256, (100, 100, 1), dtype=np.uint8) >>> >>> # Fixed blue -> yellow ramp (low == high) >>> fixed = A.Compose([A.Colorize( ... black_range=((0, 0, 255), (0, 0, 255)), ... white_range=((255, 255, 0), (255, 255, 0)), ... p=1.0, ... )]) >>> assert fixed(image=image)["image"].shape == (100, 100, 3) >>> >>> # Random thermal-ish ramp with random midpoint position >>> random_thermal = A.Compose([A.Colorize( ... black_range=((0, 0, 64), (32, 0, 192)), ... mid_range=((96, 0, 96), (160, 64, 160)), ... white_range=((220, 160, 0), (255, 220, 32)), ... mid_value_range=(96, 160), ... p=1.0, ... )]) >>> assert random_thermal(image=image)["image"].shape == (100, 100, 3)