Interesting idea, but sounds too expensive. I do believe the the image filename and alt/title attributes come into play, though those are easily spoofed so can't be taken too seriously by Google.

For example: Surrounding text, links to the page or image, topic analysis, color analysis, filename, alt attribute, OCR, advanced image recognition, Google Image Labeler (manual labeling).

Reverse image search is probably too expensive to do this for every image found on the web, but Google could surely do this for a subset. Then they could match alt attributes: If two sites use the same image, and one has an alt attribute of "automobile" and the other "car in street" you can ask Bayes how likely it is that these two sites both made fake alt text that does not describe the image.

Basically if Google Images can show you images relevant to your search term, then Google Search can find out if an image is relevant to a topic in much the same way.