The current setup is really weird and definitely wrong for many cases because it inverts the colors of the image, which is never a good idea for GUIs. So, try to OCR three different times: once on the source image, once with processing, and once with processing but no negation. This should hopefully make things work at least somewhat better for GUIs.