Request: High-accuracy ZXing alternative

380 views
Skip to first unread message

Lachezar Dobrev

unread,
Oct 2, 2012, 5:45:03 AM10/2/12
to zxing
Hello colleagues,

I am looking for alternatives to ZXing to use for a document
identification solution. ZXing is excellent to work with image
streams, where it's high speed makes it indispensable. However when it
comes to looking for bar-codes in scanned, generated or faxed-in
documents ZXing seems a blunt tool for the job.

I have received similar requests off-list (I have no idea why).

Please advise on libraries, that provide high accuracy bar-code
scanning. Both 1D as well as 2D.

I hope the ZXing team will not object my asking for competitive
products, since this is targeting a somewhat side-step use case.

Daniel Switkin

unread,
Oct 3, 2012, 10:41:01 AM10/3/12
to Lachezar Dobrev, zxing
Out of curiosity, have you experimented with the TRY_HADER flag? The intention was to do a much more thorough job of scanning, assuming infinite time and resources (e.g. server-side recognition instead of mobile). If it doesn't do the job for you I'd be interested in knowing why, and whether that's something we could fix within ZXing.

Daniel



--




Lachezar Dobrev

unread,
Oct 4, 2012, 6:04:58 AM10/4/12
to Daniel Switkin, zxing
Of course I did :)

The TRY_HARDER does indeed help to scan bar-codes, but the success
ratio is still quite low for images with the bar-code in the
upper-right corner (just for an example). I have to mention, that
sometimes people scan the documents rotated and I would like to use
the bar-code orientation to rotate the image into it's original
position.

For most documents I eventually succeed by applying various image
manipulations like crop, rescale, blur, sharpen, etc. But I've seen no
ideal combination of these, and for every image I have to do it
manually. Almost always I start with crop of the bar-code zone.
Something I would assume can be automated, but alas... At the end it
becomes easier to just input the bar-code content by hand :(

2012/10/3 Daniel Switkin <dswi...@google.com>:

Sean Owen

unread,
Oct 4, 2012, 8:08:10 AM10/4/12
to zx...@googlegroups.com, Daniel Switkin
Yeah, here you drift into machine learning land, where accuracy is important for any given image and CPU is plentiful. The mobile case is much easier since options are limited.

I can share ideas offline about how you do better detection of where the barcode is, since that is most of your problem.
I can also tell you what I know about better image processing. For still-decent images, a Canny edge detector steps up performance noticeably. For serious blur you have to be willing to make some assumptions about the barcode but then can start to apply classifier models to find barcode digits.
Reply all
Reply to author
Forward
0 new messages