WhiteRectangleDetector and high-resolution images about zxing HOT 11 CLOSED

It is already possible to configure the initial size (https://github.com/zxing/zxing/blob/master/core/src/main/java/com/google/zxing/common/detector/WhiteRectangleDetector.java#L53), and that's the right answer if you are using exceptionally large images, or just know more about how big the barcode is.

from zxing.

That is true if I am instantiating my own Detector. But DataMatrixReader instantiates its own Detector, and I don't see any way that I could tell it to use a different one. (Except by making my own Reader implementation, which would involve duplicating a lot of code.)

from zxing.

OK what about a patch to set it to a max(INIT_SIZE, some % of image size) ?

On Mon, Feb 10, 2014 at 9:23 PM, tgibson11 [email protected] wrote:

That is true if I am instantiating my own Detector. But DataMatrixReader
instantiates its own Detector, and I don't see any way that I could tell it
to use a different one.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/49#issuecomment-34684016
.

from zxing.

Shall I make this change and submit a pull request?

from zxing.

Sure open a PR. Just try running the tests first. The only issue to watch
out for is whether the change tends to make images fail that passed before.

On Mon, Feb 10, 2014 at 9:34 PM, tgibson11 [email protected] wrote:

Shall I make this change and submit a pull request?

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/49#issuecomment-34685222
.

from zxing.

This turned out to be more complicated than I first thought.

My original idea is requires the assumption that a rectangle X% of the image size is still contained inside the barcode. At least I think that is true, but you can correct me if I'm wrong. This might not be true if there is a small barcode in the center of a large image. There is also the arbitrariness of choosing a percentage.

I thought of a different approach that seems to work.

The current WhiteRectangleDetector expands the rectangle until all sides are white, which as you know might be a white area completely inside the barcode.

I propose to first expand the rectangle until each side contains at least one black point, then proceed as before.

An important assumption is that all sides of the initial rectangle are inside the barcode. I think that assumption was already being made, but it was difficult for me to be sure.

You can see my proposed changes here: https://github.com/tgibson11/zxing.

Let me know what you think.

from zxing.

Yes there is no way to know a priori where or how big the barcode is. I
think the idea is to make a guess that is likely to strike a happy medium
between working, and being fast, on images of a type that are likely to be
used.

Your assumption sounds correct and I don't see a problem with it. Of
course, you may hit the side of the image instead of something black.

Your changes don't look too hairy. If they work for your use case, and
don't cause additional images to fail the unit tests, let's commit.

On Wed, Feb 12, 2014 at 5:36 PM, tgibson11 [email protected] wrote:

This turned out to be more complicated than I first thought.

My original idea is requires the assumption that a rectangle X% of the
image size is still contained inside the barcode. At least I think that is
true, but you can correct me if I'm wrong. This might not be true if there
is a small barcode in the center of a large image. There is also the
arbitrariness of choosing a percentage.

I thought of a different approach that seems to work.

The current WhiteRectangleDetector expands the rectangle until all sides
are white, which as you know might be a white area completely inside the
barcode.

I propose to first expand the rectangle until each side contains at least
one black point, then proceed as before.

An important assumption is that all sides of the initial rectangle are
inside the barcode. I think that assumption was already being made, but it
was difficult for me to be sure.

You can see my proposed changes here: https://github.com/tgibson11/zxing.

Let me know what you think.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/49#issuecomment-34893978
.

from zxing.

Hi,

I have a similar problem with the following image

The initial rectangle is not white only, so the rectangle expand correctly at the beginning. Unfortunately, it stops at the red zone on the image below because there is a white rectangle inside the datamatrix.

Is there a way to avoid this false area detection ?

Regards

Julien

from zxing.

The idea proposed here would probably fix this. The current heuristic is "OK" but does fall down in corner cases like this.

from zxing.

Maybe I misunderstood the idea proposed by tgibson, but I don't think it will change the white rectangle problem inside the datamatrix. Once a black point is found on each border, if this "starting rectangle " is inside a bigger white rectangle, it will expand up to the white rectangle but not further.

from zxing.

Hm, yeah you may have a point there. The current process is really a simplistic heuristic that happens to work quite well for images from a camera, which are not nearly so clean or axis aligned. There's a "PURE_BARCODE" mode that's intended for this particular type of image, and that ought to work -- try http://zxing.org/w/decode?u=https%3A%2F%2Fcloud.githubusercontent.com%2Fassets%2F145699%2F16151321%2Fac423ca8-349c-11e6-8035-90bfd2a3c242.png

from zxing.