Hacker News new | past | comments | ask | show | jobs | submit login

That's interesting stuff. I wonder how the algorithms work, they have to be pretty fast if they're running on the printer itself.

Also, slightly unrelated, but this for some reason reminded me of a certain issue some copiers had a while back, where they would swap certain numbers/letters for some reason when copying. Don't really remember what the cause of that was, though.





I'd be very interested to know, as it seems to have some interesting properties: can run on limited CPU devices yet easily cope with high resolution images, in a range of colours and contrasts, and and can detect the pattern at a range of sizes and orientations.


The reasonable reason was compression. Parts of the image which looked alike were saved in the same buffer, that meant especially numbers. Which is the right way to lose the trust from your customers.


It is an interesting observation about compression loss though. The more sophisticated any compressor gets, the more its artifacts are going to look semantically plausible.

With simple technology, at first, you might get grainy or blurry images, or poor frequency response in audio. Those are easy to perceive as artifacts of the medium. Then you get things like JPEG fringes and context-dependent color shifts, which are a bit harder to learn to see past, since they are interacting with the image in sophisticated ways. JBIG (from the article) detects and compresses repeated textural elements or multiply-used glyphs, but if it applies that behavior to something that is actually information-bearing, it'll produce convincingly misleading artifacts.

There's a sequence in _A Fire upon the Deep_ where two people are conversing over a link with rapidly deteriorating bandwidth— their communicators are compressing more and more heavily to compensate, until at the end the character is "interacting" with a sophisticated Markov-like model of their interlocutor with only a few bits per second of actual entropy from the far end. Or possibly none at all.


I've run into this with cell-phones too. Someone was reading an alphanumeric value to me, and one letter reliably sounded like one number. It was really odd, but apparently some of the compression techniques for phones use speech models and it seems that the person on the other end had an accent that hit a poor spot in the model.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: