Last month I posted a call to action hoping for help with cleaning up the unpaper code, as the original author has not been updating it since 2007, and it had a few issues. While I have seen some interest in said fork and cleanup, nobody stepped up with help, so it is proceeding, albeit slowly.
What is available now in my GitHub repository is mostly cleaned up, although still not extremely more optimised than the original — I actually removed one of the “optimisations” I added since the fork: the usage of sincosf() function. As Freddie pointed out in the other post’s comments, the compiler has a better chance of optimising this itself; indeed both GCC and PathScale’s compiler optimise two sin and cos calls with the same argument into a single sincos call, which is good. And using two separate calls allows declaring the temporary used to store the results as constant.
And indeed today I started rewriting the functions so that temporaries are declared as constant as possible, and with the most limited scope as it’s applicable to theme. This was important to me for one reason: I want to try making use of OpenMP to improve its performance on modern multicore systems. Since most of the processing is applied independently to each pixel, it should be possible for many iterative cycles to be executed in parallel.
It would also be a major win in my book if the processing of input pages was feasible in parallel as well: my current scan script has to process the scanned sheets in parallel itself, calling many copies of unpaper, just to process the sheets faster (sometimes I scan tens of sheets, such as bank contracts and similar). I just wonder if it makes sense to simply start as many threads as possible, each one handling one sheet, or if that would risk to hog the scheduler.
Finally there is the problem of testing. Freddie also pointed me at the software I remembered to check the differences between two image files: pdiff — which is used by the ChromiumOS build process, by the way. Unfortunately I then remembered why I didn’t like it: it uses the FreeImage library, which bundles a number of other image format libraries, and upstream refuses to apply sane development to it.
What would be nice for this would be to either modify pdiff to use a different library – such as libav! – to access the image data, or to find or write something similar that does not require such stupidly-designed libraries.
Speaking about image formats, it would be interesting to get unpaper to support other image format beside PNM; this way you wouldn’t have to keep converting from and to the other formats when processing. One idea that Luca gave me was to make use of libav itself to handle that part: it already supports PNM, PNG, JPEG and TIFF, so it would provide most of the features it’d be needing.
In the mean time, please let me know if you like how this is doing — and remember that this blog, the post and me are Flattr enabled!