To put it another way--web browsers and web spiders are (currently) blind, and they're the ones who define what "content" is, not users. Users only deal with the web through its presentation layer; it's machines who think on the semantic level, and so it's machines who get to say what is or is not semantically meaningful.

Illustrations and charts are frequently cases of "I just explained this in prose or as tabular data; now I'll give you an alternate visual representation of the same idea as well, to enhance your understanding". A chart would work perfectly well like this:

    <img src="chart.png">
      <table>
        ...
      </table>
    </img>

Likewise, a UML diagram for an algorithm might be represented thus:

    <img src="sort_fn.png">
      <code>
        ... pseudocode for sort ...
      </code>
    </img>

Photography is basically the one special case--where the medium (the shot composition, lighting, etc.) is a large part of the message. Still, unless the photograph was taken completely at random, there is usually a reason it is being included on the page. And that reason can be used as the non-visual text. (And note that, if the entire purpose of the page is to show you the image, then you're not really looking at a hypertext document. You're more looking at HTML+CSS being used as a layout format ala PDF. There is no semantic meaning to such a page, any more than there is to a PDF.)

Here's how I think of it--you should compose every page presuming the user is blind (and deaf, etc.) first--make your page work perfectly well standing alone like that--and then perform the same process of gradual enhancement that is recommended for javascript features, when you detect that the user can deal with image/audio/video/etc. content.