Is it better to rasterize text into a header image or overlay it as text?

If you want to have a fixed header on a web page that also includes text, is it better to make that text part of the image, or to overlay the text in the header frame (such that the text is selectable, scalable, screen-reader accessible, etc)?

I'm currently in a debate over whether it would be better to embed the text in the header image on our forums. One problem is that if the background image is something we want to repeat horizontally as the page is widened, the text can't be rasterized to it because it would repeat as well - we'd need a transparent image of just the text to overlay on top of the background header image. Also, the background image selected might just fade into a solid colour for stretching, we haven't picked one out yet.

What's to be gained and lost in each scenario?

From looking at other sites on the internet, it seems like the most popular option is to rasterize the text into the header image - but a lot of times this is because the text is highly stylized as part of the logo. In our case, it's just a simple title and caption, like in a blog.

Wikipedia rasterizes:

Windows Live doesn't:

Google does for their main logo, but not for their caption: