Implement np.percentile to normalize images

[haesleinhuepf]

I recently hit a project where it would have been beneficial to do this on the GPU:

def normalize_image(image, lower_percentile, upper_percentile):
    """Make sure that image intensities lie between 0 and 1."""
    lower_bound = np.percentile(image, lower_percentile)
    upper_bound = np.percentile(image, upper_percentile)
    image_clipped = np.clip(image, lower_bound, upper_bound)
    normalized_image = (image_clipped - lower_bound) / (upper_bound - lower_bound)
    return normalized_image

For np.percentile, we could reuse our historam function. We might have to do a histogram with 10k or 100k bins and I'm not sure how efficient / doable this is.

np.clip is trivial to implement using minimum_image_and_scalar and maximum_image_and_scalar.

[StRigaud]

Hey @haesleinhuepf,

These function seems indeed fit to be implemented, do you feel like doing them? I stay around for help if needed ofc 😃

[StRigaud]

@haesleinhuepf I think we already have clip in Tier2 functions:

CLIc/clic/src/tier2/clip.cpp

Lines 10 to 19 in 80a20fa

	auto
	clip_func(const Device::Pointer & device,
	const Array::Pointer & src,
	Array::Pointer dst,
	float min_intensity,
	float max_intensity) -> Array::Pointer
	{
	auto temp = tier1::maximum_image_and_scalar_func(device, src, nullptr, min_intensity);
	return tier1::minimum_image_and_scalar_func(device, temp, dst, max_intensity);
	}

and it is already available in the python bindings