Analyzing Ballot Images of all 50 states?

First you need to get all the states to hand them over. Many seem unwilling.

As for analyzing, you don't necessarily know what you're looking for before examining them. For example - if there is a distinct pattern of markings suggesting ballot stuffing by the same person, you'd need to identify those first. Then you could train a program to identify all occurrences. But without ground truth (knowing for a fact you really spotted fraudulent ones) this wouldn't be strong proof. Absence of anomalies also wouldn't prove absence of fraud, might merely disqualify some forms of fraud.

> It seems like could settle all these conspiracy theories, disputes, and just move on.

Unfortunately, facts aren't very effective against such conspiracy theories. People who believe in them aren't (typically) engaged in a pursuit after the truth.

It's a social phenomena and has to be treated as such.

I did find this software https://github.com/FreeAndFair/OpenCount

That was funded by UC Berkley and has a tabulator part that appears to work on the common voting systems.

Readme states "Tabulator currently supports optical-scan ballots associated with Diebold (Premier), ES&S, Hart, and Sequoia ballot styles"

Since Dominion acquired Sequoia in 2014, I am guessing some small changes would make this work for the current ballot images used in 2020.

They are doing that already. When Trump kept claiming that they wouldn't let the observers see the signatures, he was neglecting to also point out that the signatures are compared electronically, and only failed signatures are reviewed by a human before getting rejected altogether. This is partly why the court dismissed this particular claim.