5. Image data types and what they mean#
In skimage, images are simply numpy arrays, which support a variety of
data types [1], i.e. “dtypes”. To avoid distorting image intensities (see
Rescaling intensity values), we assume that images use the following dtype
ranges:
Data type |
Range |
|---|---|
uint8 |
0 to 255 |
uint16 |
0 to 65535 |
uint32 |
0 to 232 - 1 |
float |
-1 to 1 or 0 to 1 |
int8 |
-128 to 127 |
int16 |
-32768 to 32767 |
int32 |
-231 to 231 - 1 |
Note that float images should be restricted to the range -1 to 1 even though the data type itself can exceed this range; all integer dtypes, on the other hand, have pixel intensities that can span the entire data type range. With a few exceptions, 64-bit (u)int images are not supported.
Functions in skimage are designed so that they accept any of these dtypes,
but, for efficiency, may return an image of a different dtype (see Output
types). If you need a particular dtype, skimage provides utility
functions that convert dtypes and properly rescale image intensities (see
Input types). You should never use astype on an image, because it
violates these assumptions about the dtype range:
>>> import numpy as np
>>> import skimage as ski
>>> image = np.arange(0, 50, 10, dtype=np.uint8)
>>> print(image.astype(float)) # These float values are out of range.
[ 0. 10. 20. 30. 40.]
>>> print(ski.util.img_as_float(image))
[ 0. 0.03921569 0.07843137 0.11764706 0.15686275]
5.1. Input types#
Although we aim to preserve the data range and type of input images, functions may support only a subset of these data-types. In such a case, the input will be converted to the required type (if possible), and a warning message printed to the log if a memory copy is needed. Type requirements should be noted in the docstrings.
The following utility functions in the main package are available to developers and users:
Function name |
Description |
|---|---|
img_as_float |
Convert to floating point (integer types become 64-bit floats) |
img_as_ubyte |
Convert to 8-bit uint. |
img_as_uint |
Convert to 16-bit uint. |
img_as_int |
Convert to 16-bit int. |
These functions convert images to the desired dtype and properly rescale their values:
>>> import skimage as ski
>>> image = np.array([0, 0.5, 1], dtype=float)
>>> ski.util.img_as_ubyte(image)
array([ 0, 128, 255], dtype=uint8)
Be careful! These conversions can result in a loss of precision, since 8 bits cannot hold the same amount of information as 64 bits:
>>> image = np.array([0, 0.5, 0.503, 1], dtype=float)
>>> ski.util.image_as_ubyte(image)
array([ 0, 128, 128, 255], dtype=uint8)
Note that skimage.util.img_as_float() will preserve the precision of floating
point types and does not automatically rescale the range of floating point inputs.
Additionally, some functions take a preserve_range argument where a range
conversion is convenient but not necessary. For example, interpolation in
skimage.transform.warp() requires an image of type float, which should have a
range in [0, 1]. So, by default, input images will be rescaled to this range.
However, in some cases, the image values represent physical measurements, such
as temperature or rainfall values, that the user does not want rescaled.
With preserve_range=True, the original range of the data will be
preserved, even though the output is a float image. Users must then ensure
this non-standard image is properly processed by downstream functions, which
may expect an image in [0, 1]. In general, unless a function has a
preserve_range=False keyword argument, floating point inputs will not
be automatically rescaled.
>>> image = ski.data.coins()
>>> image.dtype, image.min(), image.max(), image.shape
(dtype('uint8'), 1, 252, (303, 384))
>>> rescaled = ski.transform.rescale(image, 0.5)
>>> (rescaled.dtype, np.round(rescaled.min(), 4),
... np.round(rescaled.max(), 4), rescaled.shape)
(dtype('float64'), 0.0147, 0.9456, (152, 192))
>>> rescaled = ski.transform.rescale(image, 0.5, preserve_range=True)
>>> (rescaled.dtype, np.round(rescaled.min()),
... np.round(rescaled.max()), rescaled.shape
(dtype('float64'), 4.0, 241.0, (152, 192))
5.2. Output types#
The output type of a function is determined by the function author and is documented for the benefit of the user. While this requires the user to explicitly convert the output to whichever format is needed, it ensures that no unnecessary data copies take place.
A user that requires a specific type of output (e.g., for display purposes), may write:
>>> out = ski.util.img_as_uint(sobel(image))
>>> plt.imshow(out)
5.3. Working with OpenCV#
It is possible that you may need to use an image created using skimage with
OpenCV or vice versa. OpenCV image data can be accessed (without copying) in
NumPy (and, thus, in scikit-image).
OpenCV uses BGR (instead of scikit-image’s RGB) for color images, and its
dtype is uint8 by default (See Image data types and what they mean). BGR stands
for Blue Green Red.
5.3.1. Converting BGR to RGB or vice versa#
The color images in skimage and OpenCV have 3 dimensions: width, height and
color. RGB and BGR use the same color space, except the order of colors is reversed.
Note that in scikit-image we usually refer to rows and columns instead
of width and height (see Coordinate conventions).
For an image with colors along the last axis, the following instruction effectively reverses the order of the colors, leaving the rows and columns unaffected.
>>> image = image[:, :, ::-1]
5.3.2. Using an image from OpenCV with skimage#
If cv_image is an array of unsigned bytes, skimage will understand it by
default. If you prefer working with floating point images, img_as_float()
can be used to convert the image:
>>> import skimage as ski
>>> image = ski.util.img_as_float(any_opencv_image)
5.3.3. Using an image from skimage with OpenCV#
The reverse can be achieved with img_as_ubyte():
>>> import skimage as ski
>>> cv_image = ski.util.img_as_ubyte(any_skimage_image)
5.4. Image processing pipeline#
This dtype behavior allows you to string together any skimage function
without worrying about the image dtype. On the other hand, if you want to use
a custom function that requires a particular dtype, you should call one of the
dtype conversion functions (here, func1 and func2 are skimage
functions):
>>> import skimage as ski
>>> image = ski.util.img_as_float(func1(func2(image)))
>>> processed_image = custom_func(image)
Better yet, you can convert the image internally and use a simplified processing pipeline:
>>> def custom_func(image):
... image = ski.util.img_as_float(image)
... # do something
...
>>> processed_image = custom_func(func1(func2(image)))
5.5. Rescaling intensity values#
When possible, functions should avoid blindly stretching image intensities (e.g. rescaling a float image so that the min and max intensities are 0 and 1), since this can heavily distort an image. For example, if you’re looking for bright markers in dark images, there may be an image where no markers are present; stretching its input intensity to span the full range would make background noise look like markers.
Sometimes, however, you have images that should span the entire intensity
range but do not. For example, some cameras store images with 10-, 12-, or
14-bit depth per pixel. If these images are stored in an array with dtype
uint16, then the image won’t extend over the full intensity range, and thus,
would appear dimmer than it should. To correct for this, you can use the
rescale_intensity() function to rescale the image so that it uses the full
dtype range:
>>> import skimage as ski
>>> image = ski.exposure.rescale_intensity(img10bit, in_range=(0, 2**10 - 1))
Here, the in_range argument is set to the maximum range for a 10-bit image.
By default, rescale_intensity() stretches the values of in_range to match
the range of the dtype. rescale_intensity() also accepts strings as inputs
to in_range and out_range, so the example above could also be written
as:
>>> image = ski.exposure.rescale_intensity(img10bit, in_range='uint10')
5.6. Note about negative values#
People very often represent images in signed dtypes, even though they only manipulate the positive values of the image (e.g., using only 0-127 in an int8 image). For this reason, conversion functions only spread the positive values of a signed dtype over the entire range of an unsigned dtype. In other words, negative values are clipped to 0 when converting from signed to unsigned dtypes. (Negative values are preserved when converting between signed dtypes.) To prevent this clipping behavior, you should rescale your image beforehand:
>>> image = ski.exposure.rescale_intensity(img_int32, out_range=(0, 2**31 - 1))
>>> img_uint8 = ski.util.img_as_ubyte(image)
This behavior is symmetric: The values in an unsigned dtype are spread over just the positive range of a signed dtype.