Cut out objects with transparent background in the segmenter

[W7CH]

This PR modifies the output of the segmenter process by cropping objects with transparent background. The previous implementation cropped objects with an entire region of interest from the original image, which often included unnecessary background and parts of other objects. The file affected is processing/segmenter/planktoscope/segmenter/__init__.py to update and add methods for object detection, cropping and saving.

Changes:

• Define a method _save_object_image to save the object cutout with RGB channels along with the alpha channel (RGBA) for transparency
• Convert the raw image to RGB mode in the method _open_and_apply_flat
• Modify the method _stream to be compatible with RGBA mode and PNG file format of the object image
• Define a method _crop_and_apply_mask that crops the raw image and mask to the bounding box of the detected object to ensure the cutout is as close as possible to the size of the object. This method adds an alpha channel to the cropped image, and sets the alpha channel to the segmentation mask. This will make areas outside the mask completely transparent and areas inside the mask opaque.
• The nested function __augment_slice is no longer used within the method _slice_image
• For each object extracted from the raw image, the cropped mask is applied as an alpha channel to create a transparent background. The resulting object image is saved as a PNG file because the JPEG format doesn't support transparency (it ignores the alpha channel info)

Testing:

These changes were tested in an external notebook environment with multiple images to ensure that each detected object is correctly isolated and saved in the expected format. The transparency was preserved in the output PNG files (see example).

Instructions for testing this PR on the PlanktoScope

To test new changes to this PR on a PlanktoScope with internet access:

1. Push your changes in one or more commits up to the feature/segmenter branch
2. Wait for the CI checks/workflows on this PR to finish rebuilding the Docker container image for the segmenter. You should see a new entry at https://github.com/PlanktoScope/device-backend/pkgs/container/device-backend-processing-segmenter with a label (e.g. sha-55e23d1) corresponding to the latest commit (e.g. 55e23d1 ) which you had pushed to the branch. It should also have the label pr-47 attached.
3. Go to the Cockpit Terminal of the PlanktoScope (usually this will be at http://planktoscope.local/admin/cockpit/system/terminal, or else you can access Cockpit from the PlanktoScope's landing page and then you can access the terminal from Cockpit's sidebar).
4. Run the following command: forklift pallet switch github.com/PlanktoScope/pallet-standard@feature/segmentation-transparency (note: the URL given here corresponds to Bump device-pkgs to make segmentation background transparent pallet-standard#15, which deploys the latest pr-47 version of the Docker container image for the segmenter).
5. Either reboot your PlanktoScope or run: sudo systemctl start forklift-apply. This command will disconnect Cockpit in your browser tab; if you wait ~10 seconds and refresh your browser tab, then you should be able to log in to Cockpit again.

Afterwards, the updated segmenter should be running. You can check its live-updating logs with Dozzle (usually the logs will be at http://planktoscope.local/admin/dozzle/show?name=apps_ps_backend_proc-segmenter-server-1, or else you can access the live-updating logs from a link in the Node-RED dashboard's "Administration" page).

[W7CH]

A few examples of images segmented with transparent background:

Some points are not represented in the objects because they were not detected in the segmentation mask. This behavior had always existed, but segmentation with a region of interest made it possible to correct it.