You could offset the "tap the first picture" effect by e.g. randomizing the order the pictures are shown, weighting the votes of serial "top clickers" less, etc.
For example, something that should be 15-2 votes, ends up being 95-82. You'll be applying a large number of votes to both sides, and pushing everything towards a 50/50 rating. This doesn't help anyone, the goal is A/B testing, and you're making it more difficult to get accurate data. 15-2 shows a lot of promise for A, but then you add 80 random votes to both sides, and 95-82 seems like a tie.
It's all in the presentation. If you just highlight one image and stamp "WINNER" next to it, most people won't even look at the numbers. Crowning a winner is more important than being scientifically accurate.
I suppose the question then becomes: will the end user notice or care if the votes are random? Do the votes need to come from humans at all?
Instead of showing the user 95-82, or 25-12 you tell the user "others prefered shirt A 2-to-1 or shirt b 66% to 33% or whatever might be appropriate.
Anyway, again, I'n not suggesting any particular techniques are the right ones, just that there (almost certainly) viable techniques available to mitigate the problem.