That probably is doable on modern local hardware, but on network disks, it might still be slow.
Make that an IMHO not unreasonably large 500 photos and your local disk might be slow, too.
Add a virus scanner, and 100 photos might already be annoyingly slow.
⇒ I would try to find a way to store away the checksum somewhere in the file (say in an extended attribute, as a (modification timestamp, checksum) pair)
If you do that, you only have to read the whole image the first time. Once you created the thumbnails, you only need to read them, and they’re a lot smaller, and can be put in a single file.
The proposed approach reads a significant part of each file to determine whether the cached thumbnail is still current. That doesn’t give a huge improvement over not caching thumbnails at all (factor of 5 in bytes read/network traffic, as a rough guess, likely less in disk seeks)
FM that takes a lot of memory or FM that use computation resources when required only. I tend to think that in general expectation from FM is to be very small in memory demands and take valuable computational resources only if it is absolutely necessary.