The current file de-duplication system uses things like file timestamp and MD5 hash of the content. It would be helpful if an option to detect visually similar images was added too.
There are some open source projects that do this already, maybe useful as a reference or an open source library or something.
https://dupeguru.voltaicideas.net/
https://ermig1979.github.io/AntiDupl/english/index.html