just curiosity: AI image analysis ?

[Michael Miller]

Hi everyone,
I’m Mike, the digiKam developer responsible for the AI in digiKam.

I really like the idea of a pluggable architecture to consume online services for object identification and image classification.  On the digiKam team, we take privacy very seriously which is why we haven’t incorporated many external cloud services into digiKam (Google Maps and online translation being the exceptions).  We believe your data and images are yours, and we don’t want to do anything that could cause a user to unknowingly share their data and images.

Having said, that I think there are ways we could safeguard user data and still allow a mechanism to consume cloud services.  The AI/Machine Learning pipelines in digiKam are easily extended, and we could add a generic pipeline that calls a user configurable command in a shell.  As long as the command returns a value or values back to digiKam in an agreed upon format, the end user could create their own scripts to call cloud services.

Since the user would have to create or install the scripts themselves, I think it fair to assume the user knows the data is being shared with a cloud service.

I’ll talk with Gilles and Maik and get their opinion.  I’m just one person on a talented team of developers, and I’d like to get their thoughts.

Cheers,
Mike

> On Mar 31, 2025, at 4:29 PM, Jens Benecke <jens-digikam at spamfreemail.de> wrote:
>
> Hello,
>
> I think this goes in the direction of something I proposed a few years ago.
>
> Situation:
>
> - There are a lot of online services offering specialized AI based $WHATEVER. Tagging, background removal, you name it.
> - Most of these have an API where you can call some HTTP endpoint and either upload a (possibly reduced) version of your image, or a part of it.
> - They then - usually - return an image or a set of text based data, like tags or free text.
>
> Digikam needs a system that can use these APIs in a flexible way, e.g. by incorporating scripts directly integrated into the app. There is a scripting interface but it cannot feed data back into Digikam or replace images.
>
> https://bugs.kde.org/show_bug.cgi?id=384444
>
> At the time, this idea was rejected for privacy reasons.
> But IMHO, if you want to do specialized tagging using an online service, and remain API agnostic within Digikam, this is not necessarily a problem.
>
> Maybe reconsider?
>
> Regards,
> Jens
>
>
> Am 31.03.25 um 21:49 schrieb support at hausoos.com:
>> Yes, of course! But if we had a way to integrate specialised image recognition, running locally or through an API on connection, it would be interesting for some people (of course, niche groups, but still ....).For example, Plantnet has an API which can connect to the AI recognition engine. Plantnet can reach > 85% accuracy with good set of pictures for example (that varies across taxa, and across geographies).
>>
>> There are quite a few other models, such as Flora Incognita, which are very good for the plants in their database and have versions of te models weighted that can be downloaded and run locally.
>>
>> That said, it would be very interesting to see whether it is possible to at least identify down to the genus level, even if not the species.
>>
>> The same could be done for other categories (trains, for example, for those who are passionate about them, cars, buildings, fossils, etc. .... :-) ....)
>>
>> Kind Regards
>>
>> Corrado
>>
>> Quoting Remco Viëtor <remco.vietor at wanadoo.fr>:
>>
>>> On dimanche 30 mars 2025 15:45:00 heure d’été d’Europe centrale Bill Allen
>>> wrote:
>>>> I did a couple of tests using the python API in which I fed my ex if
>>>> metadata as part of the prompt which helped, but it's not going to work
>>>> like, say, plant net for identification of plants.
>>>
>>> There are two big differences between systems like plantnet and digikam:
>>> - in Digikam, the whole model is *local*, systems like plantnet can have many
>>> more resources available;
>>> - Plantnet is a *specialised* system, Digikam would need a general system
>>>
>>> Note that Plantnet, while good, can give wrong results. Other such systems can
>>> do a lot worse (some specialised systems barely reach 50% success, and those
>>> are the good ones *for their domain*).
>>>
>>> Remco
>>
>>
>>
> --
> Regards, Jens
>
>