Well, cleaning up the image doesn't really belong in the bar-code scanning program, anyway (IMHO). We can use netpbm; something like what we want may be gotten with pamtilt & pnmrotate (to auto-straighten), maybe pnmnlfilt (to clean up noise), and pnmnorm & pamdepth (crank up the contrast and squash to one bpp).

Then you've got a pretty good image, but maybe you still don't trust just one row out of the middle, but would rather an average? OK, pamcut out an arbitrary chunk, say the center 1/2 vertically, or maybe try to guess heuristically. Then pamscale -nomix can be used to squash it down to a single row; it'll take the more common color for each column.

Now you've got a single row of pixels, ready to apply that algorithm as it stands. That's what I'm thinking, anyway; if that's too slow, then the slow parts of the image processing can be converted to dedicated C code.

But I think scripting it together should give decent performance. I'll play around with it some; it'll probably be a couple weeks before I can look into porting it, but I'll play with the pbm pipeline, and with your code in Octave.