How about, if you'd take an 8 bit grayscale image, and allow each pixel a margin of error, say 2 values up or down. Then you could process the pixels sequentially while looking for the next value that is within the range of +/- 2 of your next BF token and adjust it to the correct value. If instead you happen upon an incorrect BF token, you can also adjust it so that it either turns into the correct one or a non-BF instruction byte.
Of course there's no guarantee that you'll be able to fit the whole program since you'll be skipping a lot of pixels. But an image is big and the message is short, so there's a good chance. Especially if you retry with different BF encodings of the same message, and failing that upping the error margin to +/- 3, then +/- 4, etc.