Apart from the cultural neutrality of the image, you’d also need to find words that have completely different spellings in each language and images that are precise enough not to create too many options for a user’s response (for instance, in English, I could see “truck,” “jeep,” and “SUV” all being possible responses to an image of a truck). As with question-based CAPTCHAs it needs to be dead simple and should not leave itself open to the user over-thinking a response (e.g. “How much is 1 plus 1?” or “Is ice hot or cold?”).

It is a really cool idea though; I’d love to see a prototype.