Pixel-level processing

Faced with an image scanned in to the computer, the first group of processes deal with pixel-level transformations.

• Thresholding is the determination of whether a particular pixel position is to be treated as white or black, given that it is actually perceived as some level of color or gray: We don't address this task in our programs, relying instead on the scanner (perhaps with its low-level software) appropriately adjusted, to come up with the (binary level) image. While such an arrangement can be fooled by inverse-color printing, printing on top of halftone, etc., we have found that the scanners we've used can be adjusted satisfactorily to produce 2-level images. This decision to leave well enough alone could be re-examined and we could either write our own thresholding program or directly use a gray-scale. This latter approach would seem to be far most costly that our binary bitmap approach, but it seems plausible to trade-off some low-resolution gray-scale for high-resolution 2-level images.
• Noise reduction. This can include a host of transformations attempting to modify or filter the shapes represented, including morphological processing and ``kFill'' filters. We don't do this, although it could be added. We do not provide this; partly we have not found it very useful; Further, at least some of these processes are not entirely natural in our representation.

• Thinning/skeletonization This is a kind of higher-level morphological concept that can be applied to images which is especially useful in images that are graphs, maps, etc. Thinning for the purpose of text recognition appears less appropriate. A solid circle thins to a dot. A slice of swiss cheese is unrecognizable.

• Chain coding and vectorization We do not do either of these per se, though run-length encoding on a row-by-row basis serves some of the same needs: it is easier to compute connectivity, and it is potentially far more compact.

• Region detection/ connected components