OpenCV dependency removal

[PavlosMelissinos]

I'm interested in throwing away every import cv2 line because OpenCV has mostly been a pain for no reason. The attempt can be seen in the no_opencv branch

The strongest contenders are pillow and scikit-image.

I will be documenting my experiences here mostly as notes to myself and possibly to fuel a discussion.

Library	Wraps NumPy	channel order	dimension order
OpenCV	yes	bgr	(width, height)
Pillow	no	rgb	(width, height)
skimage	yes	rgb	(height, width)

Any suggestions are welcome!

[PavlosMelissinos]

I'm going with pillow after all, as the conversion of a PIL.Image to numpy is trivial and pretty much contains every necessary operation. Next few commits will gradually simplify preprocessing (currently a flag determines which library will be used and that unnecessarily increases code complexity).

[PavlosMelissinos]

CLAHE is one of the things that does not seem to be supported by pillow. This is not a deal breaker but is anyone aware of an alternative implementation? I know there's equalize in ImageOps but I don't think it uses CLAHE (which works great by the way, I will upload a notebook with a demonstration)

[PavlosMelissinos]

Here is the notebook, unfortunately I couldn't get ipywidgets to show the outputs at all in the static version.

[mrgloom]

Can you explain why opencv was dropped?

[PavlosMelissinos]

Hey, sorry for taking this long but I haven't been around in general this past year.

It's a personal choice. I prefer lightweight, simple options that stay away from obscurity and opencv is neither. It follows a different approach than most libraries (e.g. bgr vs rgb), sometimes for no good reason, and this project didn't need any of the advanced features of opencv, therefore I decided against using it. Of course as this is an open source project, anyone is free to modify the code and add it. :)