Moving `imageops` and similar functionality to the `imageprocs` crate

[ripytide]

Tracking Table

Key

• 🟥: Not yet in imageproc
• 🟧: In imageproc but not yet released.
• 🟪: Released in imageproc but still in image
• 🟩: Deprecated in image

function	status	equivalent in imageproc
blur	🟪	filter::gaussian_blur_f32
crop	🟧	compose::crop (slightly different in that it returns an entirely new image vs returning a SubImage that does no processing)
crop_imm	🟥	cannot be moved as it returns a SubImage that does no processing, compose::crop can be used instead but returns a new image
filter3x3	🟧	filter::filter or filter::filter_clamped
flip_horizontal	🟧	compose::flip_horizontal
flip_horizontal_in	🟥	None
flip_horizontal_in_place	🟧	compose::flip_horizontal_mut
flip_vertical	🟧	compose::flip_vertical
flip_vertical_in	🟥	None
flip_vertical_in_place	🟧	compose::flip_vertical_mut
horizontal_gradient	🟥	None
interpolate_bilinear	🟧	Implemented but private and needs refactoring relevant issue
interpolate_nearest	🟧	Implemented but private and needs refactoring relevant issue
overlay	🟧	compose::overlay
overlay_bounds	🟥	Only makes sense as a helper function for overlay and replace so I don't see the point migrating this a public function
replace	🟧	compose::replace
resize	🟥	None
rotate90	🟧	geometric_transformations::rotate90
rotate90_in	🟥	None
rotate180	🟧	geometric_transformations::rotate180
rotate180_in	🟥	None
rotate180_in_place	🟧	geometric_transformations::rotate180_mut
rotate270	🟧	geometric_transformations::rotate279
rotate270_in	🟥	None
sample_bilinear	🟥	None
sample_nearest	🟥	None
thumbnail	🟥	None
tile	🟥	None
unsharpen	🟥	Nearly, filter::sharpen_gaussian exists but lacks the threshold parameter and doesn't work on color images relevant issue
vertical_gradient	🟥	None
colorops::brighten	🟥	None
colorops::brighten_in_place	🟥	None
colorops::contrast	🟧	contrast::stretch_contrast
colorops::contrast_in_place	🟧	contrast::stretch_contrast_mut
colorops::dither	🟥	None
colorops::grayscale	🟥	None
colorops::grayscale_alpha	🟥	None
colorops::grayscale_with_type	🟥	None
colorops::grayscale_with_type_alpha	🟥	None
colorops::huerotate	🟥	None
colorops::huerotate_in_place	🟥	None
colorops::index_colors	🟥	None
colorops::invert	🟥	None

Original Comment

At the moment the image crate has several image processing functions in
the imageops module which seems a bit strange to me as they look like
they would better belong in the imageproc crate. Is there a reason why
such functionality is in this crate, or is perhaps a holdover from long
ago?

I would be happy to do the work migrating/integrating the functionality
into imageproc if you'd like.

[fintelia]

The imageops module was added long before my time and has gotten very little attention in the last few years. It contains some methods like crop or resize that totally make belong in the crate, but also has others like huerotate that probably aren't necessary. I'd be open to moving functionality, but we should discuss on a case-by-case basis first

[ripytide]

Great, image-rs/imageproc#7 is the corresponding issue on the imageprocs crate.

My current action plan is to write up a list of all the functions in imageops and then individually integrate each one into imageprocs. After each function is integrated and released that would give you the ability to deprecate those individual functions and point users at the corresponding functions in the imageprocs crate. Then at some point in the future you would have an easier time deleting the deprecated functions hoping most users will have migrated.

How does that sound?

[theotherphil]

As the image crate has a huge number of users and very little activity happens on the imageops code, the cost of maintaining duplicated functionality is low but the impact of breaking changes is potentially high.

I'm in favour of:

1. Adding all current imageops functionality to imageproc, either vs re-exports or re-implementations depending on whether we want to tweak anything.
2. Adding a note in the imageops module (via code comments/readme/PR template/whatever) that new image processing functionality belongs in imageproc.

The question of deprecation or retiring existing imageops functionality is trickier and thankfully not up to me!

(Edit to add context for other readers: I’m the maintainer of the imageproc crate.)

[fintelia]

Yeah, this would involve deprecating the methods now and only removing in the next major release (which won't be any time soon given we haven't even started planning it yet)

During this process it is also worth looking at whether these methods are using internal functionality / are tightly tied to the image crate. Like the invert method only works because one of the required methods on the Pixel trait is Pixel::invert. The APIs around writing image processing code that's generic across image type is something I've held off working on because it seemed like there were no elegant options and any changes would just cause a lot of breakage

[ripytide]

I've written up the list in a tracking table in the top comment which can be updated as we make progress toward this larger goal. Let me know if you'd like to change anything or if I've missed anything.

[fintelia]

To clarify, the last step should probably be deprecated from image crate. We'll remove all the depreciated items in one batch when we do the 0.26 release in 6-12 months (or whenever it happens... haven't started planning that release yet)

[ripytide]

I've finished my case-by-case analysis of the current state of things and it doesn't look great, 1/43 functions has an equivalent version in imageproc. This might be more work than I anticipated 😄

[johannesvollmer]

As the image crate has a huge number of users and very little activity happens on the imageops code, the cost of maintaining duplicated functionality is low but the impact of breaking changes is potentially high.

This is true, but removing that functionality still has great value for simplifications across the library. For example, the Pixel::invert functionality is only required for image processing, but it is not helpful when trying to integrate non-standard decoders for niche pixel formats.