Face recognition AI / DL and adding features

[Gilles Caulier]

Le lun. 17 août 2020 à 20:26, <cerp at eeos.biz> a écrit :
>
> Dear Gilles,
>
> Thanks a lot. Some comments in the text.
>
> Quoting Gilles Caulier <caulier.gilles at gmail.com>:
>
> > Le lun. 17 août 2020 à 13:34, <cerp at eeos.biz> a écrit :
> >>
> >> Dear Gilles,
> >>
> >> Thanks for the link. I looked at the bugzilla thread, and it is very useful.
> >>
> >> In my case, I think we may need to start developing a DL classifier
> >> earlier, I cannot wait for Digikam to have stabilised the face search,
> >> so I may have to go a different way if I am not allowed to create a
> >> Digikam plugin for plants classification. We have several 10s of
> >> thousands of pictures, with many TB of pictures, and manual screening
> >> is now impossible. I am happy to share my results of course.
> >
> > About plants we have this entry to create a new plugin :
> > https://bugs.kde.org/show_bug.cgi?id=394544
>
> Thank you! This is exceptionally useful.
>
> >
> >
> >>
> >> Thanks for the link to Photils, which seems to be useful for us. I am
> >> not completely clear: is photils the engine that is currently used in
> >> Digikam for face recognition with the new AI / DL approach, or is this
> >> in the future?
>
> I looked at Photils and tried it. I think it may be useful but not
> really optimal for our use. We need something that we can train at
> different levels of detail (initially plant /  not plan, then family,
> species, subspecies etc. etc.) , possibly based on Deep Learning.

I don't yet tried Photils as well. the plan is to make a Generic
plugin for digiKam which will use it and propose keywords to tag
images.

>
> >
> > No. digiKam as a dedicated internal engine to process detection and
> > recognition based on AI.
>
> What technology does Digicam use? I assume Digikam is using DL, but
> what starting from which libraries? I assume you did not develop
> everything from the ground up, because there are so many interesting
> libraries available.

Of course, we don't re-invent the wheel. Take a look to the 7.0.0,
there are all technical details, history, and plan explained :

https://www.digikam.org/news/2020-07-19-7.0.0_release_announcement/

>
> >
> > Photils is an alternative which compute a fingerprint of images, send
> > it to a AI engine which try to identify objects, monuments, animals,
> > etc... It return a list of keywords.
>
> I understood the mechanism, and trialled it, and it turns out to be
> too generic for our use. I would prefer something we can train, and
> that can be used at different level of details (initially plant /  not
> plant, then family, species, subspecies etc. etc.).

Yes. I talk with the student working on face engine AI based, and this
plan are possible but complex. The problem is the model of data for
the deep learning engine. 2 solutions : use a pre computed model (the
case currently for the face) or compute the model step by step
depending of the king of form to detect and recognize. This last case
is really more complex and long to implement, test, and validate.

Regards

Gilles Caulier