I researched for a long time how I might have a 10000 image system. The challenge is that one would want the images to be sufficiently different from each other; and for the numeric value of the image to ideally be hinted at by the nature of the image.
So I looked at how a sequence of 100 images might all be under a topic such as Birds. But it is a big challenge to find 100 bird images all unlike each other to avoid confusion.
Also, if I find what looks like a feee image on the internet, what if I later learn that it is copyrighted and has rules about its use?
So I started looking at people faces which could be generated by software - and the type of head could be a clue as to the numeric value of the person within a 10000 system of images.
The other aim was to use the Town 100 syllables (also in the People from 00 to 99) to make names for these people. So, since 35 is KO and 39 is KU, a person named KOKU would be representing 3539 and also the possible foreign language word word 'KOKU'. However, would the time put in to making and learning the system be justified?
A system of about 5000 images seemed more practical. I can have a rule that 3539 represents either KOKU or a value where the leading digits of KO are switched with the syllable offset 50 from it: 35+50 = 85. 85 is WO. So the word WOKU can be represented by the same image.
If I have a story where one of these people does an action on an object, the action or the object can be a clue to add 5000 so that I translate the 5000 system image as either its usual number or as 'that number plus 5000'.
The 5000 images will likely be made from 1000 people head shapes but the hair and other attributes can make the people significantly distinguishable from each other.