Pipeline-runnable code to "decollage" the combined FlowCam images

[metazool]

This code currently exists as a command-line script repurposed by @Kzra (based on https://sarigiering.co/posts/extract-individual-particle-images-from-flowcam/) and run by hand on a VM with internal storage available as a volume mount

#20 (Automating file transfer off the FlowCam) suggests we may be able to change this workflow beneficially and take the VM out of the loop.

This issue is about setup work to add the code with tests to the decollage package and also to preserve more metadata while doing it - step towards replacing #4 (minimal metadata, currently just a file listing!) and recording more detail (coordinates, date, depth, and also image size)

Not end-to-end as per #11 (diagrams of the instrument-to-storage workflow) but could be a chance to get a handle on the Luigi package in passing (a python analogue to R's {targets} recommended by @albags - rapid prototyping for work with an Airflow destination?)

[metazool]

See also https://github.com/NERC-CEH/cyto-ML/issues/6 (private to internal contributors)

• comment here about embedding the metadata in the output image EXIF headers which makes a lot of sense here, the files are still in transit to cloud storage where they'll get fixed addresses, it saves rewriting external metadata

And is better for compatibility with scivision if that's still actively being supported by Turing Inst (who does one ask about that?)

[metazool]

This is interesting! The decollage script depends on the presence of a .lst file (which we have available) - it only uses the pixel coordinates to fish the image out, but all the geometry analytics based on the binary mask is included in this file too

(screen photo from the FlowCam walkthrough about 2/3 of the way down:

it's an "almost but not quite CSV" with 53 lines of column_name|type headers and there's bound to be a lot we can get out of these metrics alone (I collaborated on a paper about this once, characterisation of grains in sandstone, comparing shape-based to deep learning approaches but it got boiled down to "does the deep learning approach work" rather than digging into the tradeoffs between the approaches...)

[metazool]

Glad to close this, there's a next step preserving the analytic metadata in the .lst files off the FlowCam but that's out of scope here