I wanted to try for a long time and finally stumbled upon https://github.com/lucko515/search-book-by-cover-server.



It seems abandoned but basically is trying to match book covers to a dataset of book covers.
It's a https://en.wikipedia.org/wiki/Content-based_image_retrieval system.

That could easily translate into "match LD covers to a dataset of LD covers" -- and a dataset of LD cover is something we have: 43,873 covers to be precise.

So, after fixing a few typos, wrong filenames and python packages dependencies, I am currently deep learning all the covers.
It takes about 24 hours.



However I tried on a subset of the covers before and the results were... not good

Let's see with the full dataset.

Julien

Oh well, process crashed at 96%|(42027/43873) [24:12:59] on a missing cover file but we can use the index up to 42000.

The matching takes an awful lot of time (the index is ~2.9GB).

And... it can't even match itself? (was my mistake)

OK... back to the drawing board.

Julien

Tried an easy match with a cover that has not yet been added for testing:



Quite a particular color/shape that should find easy matches.

Loading/analysing the dataset dictionnary takes 130s (more than 2min.) per match. That's not usable in real life.

First 3 are good:



But then it gets weird:



And it completely missed:



I would imagine the heavily colored obis are throwing the RGB histograms completely off, they would have to be removed (masked -- luckily we have been describing the obi location on the cover for years!) from the learning phase. The top banner "widescreen" or "dts" is probably also creating comparison data that put the matching off-rail. Color grading from cover to cover will also be a problem if the cover was shot under natural or artificial lights.

Julien

If this ever works, it will require some careful masking to remove all common elements.

Example, trying to match the French PAL Robocop 2 release for other Robocop 2 covers...



... fails to find other Robocop2 but succeeds in matching the common right vertical part:




Julien

Same issue with other cover that share a lot of visual elements.

Example: The Music Disc collection

This cover:



Will match similar designs:




and for some reason this one as well:



Oh well, it was an interesting experience but I don't see how it can provide any meaningful added value to the website.

I'm giving up!

Julien

I gave it one last round with a full indexation:



But the matching is... what you would get from aiming at a target in the dark.
Totally unable to perform matching if the source material isn't perfect.

I'll revisit the idea someday if I find a better solution.

Google Images does it better, and in less than 1 second.

Julien