PageSegMode.OsdOnly orientation always PageUp

[boddy33]

Hi,

I need to detect page orientation only (no OCR).

If I process image with PageSegMode.OsdOnly, detected Orientation is always PageUp:

using (Bitmap src = Image.FromStream(imageStream) as Bitmap)
{
using (TesseractEngine tesseractEngine = new TesseractEngine(_tessdataPath, "eng", EngineMode.Default))
{
tesseractEngine.DefaultPageSegMode = PageSegMode.OsdOnly;

                        using (var page = tesseractEngine.Process(src))
                        {
                            using (var iter = page.AnalyseLayout())
                            {
                                iter.Begin();
                                var pageProps = iter.GetProperties();
                                Orientation orient = pageProps.Orientation;

...
}}}}

this code seems to work only if DefaultPageSegMode = PageSegMode.AutoOsd;

Any idea why?

Thanks
Bogdan

[charlesw]

Do you have the orientation scripts installed in your tessdata directory, check out Issue #109.

[boddy33]

Yes, osd.traineddata is included and PageSegMode.AutoOsd works correctly. PageSegMode.OsdOnly always brings PageUp

I am trying to get PageSegMode.OsdOnly work because I think it will work faster than PageSegMode.AutoOsd: In our system we firstly detect orientation, rotate file and then process page with PageSegMode.SingleBlock. We do processing with multiple threads and trying to reduce processing times.

Regards
Bogdan

[charlesw]

Thanks, I've had a look into this and apparently your supposed to use the
DetectOS function if you're just after orientation information OsdOnly).
Unfortunately that functionality hasn't yet been exposed via the CAPI that
this wrapper uses. I'll see if I can get it added in the upcoming tesseract
3.04 release but no promises :)
On 26 Feb 2015 06:27, "boddy33" notifications@github.com wrote:

Yes, osd.traineddata is included and PageSegMode.AutoOsd works correctly.
PageSegMode.OsdOnly always brings PageUp

I am trying to get PageSegMode.OsdOnly work because I think it will work
faster than PageSegMode.AutoOsd: In our system we firstly detect
orientation, rotate file and then process page with
PageSegMode.SingleBlock. We do processing with multiple threads and trying
to reduce processing times.

Regards
Bogdan

—
Reply to this email directly or view it on GitHub
#156 (comment).

[charlesw]

I've had another look into this and I was wrong when I mentioned that this functionality was not exposed via the CAPI 3.02. I should be able to expose this functionality however it should be noted that it only returns the orientation (e.g. Page UP, Page Down, etc) and script. I'm not sure I can expose the script information as this generally requires access to some C++ classes which I don't have access to from .NET. Note the deskew angle is NOT exposed\available. If you need that information you'll need to use AutoOSD with AnalyseLayout.

[boddy33]

Hi Charles,

thank you for information.
We are also running some tests with 3.04 and noted that OSD quality is much higher than in 3.02 (especially for low quality faxed documents) so we have upcoming in focus.

Can we contribute somehow in speeding up 3.04 release?

[charlesw]

I've implemented this support in the current dev branch (targets 3.02) and it will be in the next release, probably latter this week. Can you also test it to see if it's working for you? Note that it only exposes the orientation and corresponding confidence not the script information as I couldn't get hold of that easily.

In regards to supporting 3.04 the main blocker here is that this has to wait until after 3.04 has been officially released. After that it shouldn't take to long to release an updated version of this wrapper that uses this version of Tesseract though it probably won't expose ALL of the new functionality right away (I tend to weight until someone explicitly requests something, ideally with a particular use case in mind).

Usage Example:

using (var img = Pix.LoadFromFile(@".\phototest.tif")) {
    using (var page = engine.Process(img, pageSegMode)) {
        Orientation orientation;
        float confidence;
        page.DetectBestOrientation(out orientation, out confidence);

        Assert.That(orientation, Is.EqualTo(Orientation.PageDown));
    }
}

[boddy33]

Hi Charles,

thanks a lot for this improvement, it works. I will test a bit more and let you know results.

[ghost]

Hi,

It will be very helpful in a specific project that I´m working. Congratulations!

Carlos Flávio Barreto Ferreira de Souza

[boddy33]

Hi,

so according to our tests accuracy of orientation is about 95% with DetectBestOrientation in 3.02 comparing to 98% with custom 3.04 build. But 3.04 is 35% slower than 3.02.

Still cant understand though why OCR with AutoOSD brings such a poor results comparing to DetectBestOrientation in same 3.02 build.

Waiting for official release of wrapper, thanks again Charles.

[charlesw]

Okay, thanks for the extra testing. I'll do an official release later today.
On Mar 29, 2015 7:00 AM, "boddy33" notifications@github.com wrote:

Hi,

so according to our tests accuracy of orientation is about 95% with
DetectBestOrientation in 3.02 comparing to 98% with custom 3.04 build. But
3.04 is 35% slower than 3.02.

Still cant understand though why OCR with AutoOSD brings such a poor
results comparing to DetectBestOrientation in same 3.02 build.

Waiting for official release of wrapper, thanks again Charles.

—
Reply to this email directly or view it on GitHub
#156 (comment).

[charlesw]

The new release 2.3.0 is out.