pil2cv: allow (and drop) alpha channels on input imgs

[bertsky]

The error I got by calling cvtColor as it was:

OpenCV(4.8.0) /io/opencv/modules/imgproc/src/color.simd_helpers.hpp:92: error: (-2:Unspecified error) in function 'cv::impl::{anonymous}::CvtHelper<VScn, VDcn, VDepth, sizePolicy>::CvtHelper(cv::InputArray, cv::OutputArray, int) [with VScn = cv::impl::{anonymous}::Set<1>; VDcn = cv::impl::{anonymous}::Set<3, 4>; VDepth = cv::impl::{anonymous}::Set<0, 2, 5>; cv::impl::{anonymous}::SizePolicy sizePolicy = cv::impl::<unnamed>::NONE; cv::InputArray = const cv::_InputArray&; cv::OutputArray = const cv::_OutputArray&]'
> Invalid number of channels in input image:
>     'VScn::contains(scn)'
> where
>     'scn' is 2

Removing the alpha channel in advance avoids the problem.

But perhaps we should not naively drop the channel, but use it (somehow) to mask the image? (I.e. for segmentation: no layout analysis in fully transparent zones, for binarization: pure background)

[vahidrezanezhad]

@bertsky @kba @cneud Even without any changes, including this PR, Eynollah (which I tested using my latest PR ) works correctly on this image with an alpha channel
. It seems this issue happens with other images with alpha channel. can you provide me images with alpha channel that this issue happens @bertsky ?

[kba]

If we decide to move ahead with this, we should backport that to OCR-D/core whence I took it originally.

[bertsky]

Ok, so to revisit this...

The image @vahidrezanezhad used, which is data/gutachten/data/IMG/IMG_1.tif from OCR-D's test assets repo, actually is a two-page TIFF with a simple RGB up front (which is why it is well-behaved regarding OCR-D and Eynollah) and an RGBA (but actually RGBa, i.e. premultiplied) as second layer. Extracting and running with the latter image in lieu of the original one does work, but produces bad results due to the wrong image metadata (RGBA instead of RGBa):

But depending on the base layer (gray or palette or rgb), the OCR-D interface (i.e. pil2cv()) can still fail with the above stacktrace. As it does for this example image.

Hence my latest commit on #222 as a stripped down version of this PR.

The standalone CLI uses cv2.imread with default flag cv2.IMREAD_COLOR_BGR, which does work in many cases already.

I am not so sure anymore what my original example was, but will close for now.