I'm fairly certain it actually is classifying image contents since at least a year or two ago (or maybe a little bit even longer – I can't remember when exactly I first noticed this), because I've noticed getting image search results that can't be explained otherwise (the keyword I've been searching for most certainly doesn't appear anywhere on the page containing the image in question). It can be turned off by doing a verbatim search (i.e. surrounding the search term with quotes), but otherwise it definitively happens, and it's occasionally been quite useful actually.

Likewise, since at least around the same time (if not somewhat earlier), Google has also been running OCR on all images it indexes.