Problem with png images

[Dian8]

Environment

• Tesseract Version: tesseract-4.0.0-beta.4 (binary from github releases, dependencies from cppan)
• Platform: Windows 10, x64

Sample code:

        Pix *image = pixReadMem(buffer, length);

        api->SetImage(image);

        //pixWrite("c:\\temp\\img.jpg", image, 2);
        //Pix *image2 = pixRead("c:\\temp\\img.jpg");
        //api->SetImage(image2);

        api->SetSourceResolution(70);

        // Get OCR result
        char* outText = api->GetUTF8Text();
        printf("OCR output:\n%s", outText);

Current Behavior:

If source image comes in jpg all works good. If image comes in png recognized text is empty.
If convert source png to jpg (commented code fragment above) text is recognized well

Expected Behavior:

Png images successfully recognized

[amitdo]

Try the command line and post the output.

Can you share the png image?

[zdenop]

Can you also provide output of tesseract -v?

[Dian8]

tesseract.exe -v

tesseract 4.0.0-beta.3
leptonica-1.76.0 (Sep 18 2018, 16:25:33) [MSC v.1914 LIB Release x64]
libgif 5.1.4 : libjpeg 9b : libpng 1.6.35 : libtiff 4.0.9 : zlib 1.2.11 : libwebp 0.6.1 : libopenjp2 2.3.0
Found AVX
Found SSE

Png image is

[Dian8]

Any suggestions?

[zdenop]

Remove alpha channel from png image.

[zdenop]

@amitdo @stweil @jbreiden : one of solution (workaround?) would be to remove alpha channel in tesseract. e.g. during SetImage:

index 9d0a5a51..a156c1ee 100644
--- a/src/ccmain/thresholder.cpp
+++ b/src/ccmain/thresholder.cpp
@@ -150,6 +150,14 @@ void ImageThresholder::SetImage(const Pix* pix) {
   if (pix_ != nullptr)
     pixDestroy(&pix_);
   Pix* src = const_cast<Pix*>(pix);
+  int spp = pixGetSpp(src);
+  // Remove alpha channel
+  if (spp == 4) {
+    PIX* tmp = pixRemoveAlpha(src);
+    pixSetSpp(tmp, 3);
+    src = pixCopy(nullptr, tmp);
+    pixDestroy(&tmp);
+  }
   int depth;

Not sure if we have time to check this for side effect.

[zdenop]

remark: pdfrenderer.cpp also removes alpha channel from PNG images with pixAlphaBlendUniform function...
So it seem to make sense to get rid of alpha channel at the beginning....

[zdenop]

Alpha channel seem to be only part of problem. I exported image with transparency to png and gif. Tesseract produce:

>tesseract.exe i1914r.png -
Empty page!!
Empty page!!

>tesseract.exe i1914r.gif -
Warning. Invalid resolution 0 dpi. Using 70 instead.
Estimating resolution as 244
m 130000

=> for png it is not able to do recognition. But when I use psm:

f:\Project-Personal\tesseract.test>tesseract.exe i1914r.png - --psm 6
S
m 130000

f:\Project-Personal\tesseract.test>tesseract.exe i1914r.gif - --psm 6
Warning. Invalid resolution 0 dpi. Using 70 instead.
m 130000

it produce output but worse than for gif... So more investigation should be done. Here are images for testing.

i1914r.gif i1914r.png

[amitdo]

Zdenko, did you try to do what the pdf renderer originally did? Your code uses a different function.

[zdenop]

Explained in code review comment...

[stweil]

Removing the alpha channel does not change the OCR result for this test image. There is an alpha channel, but there are no transparent parts in the image, so that's not surprising.

The real One problem of the original test image is the resolution: it is missing, and Tesseract guesses a resolution of 147 dpi for the PNG image. With a resolution of at least 150 dpi Tesseract recognizes text "9) 130000". The same is true for other image formats. JPEG only works because Tesseract guesses 244 dpi for that format.

So I see two issues here:

• Why does Tesseract guess different values for the resolution for the same image when only the image format is changed?

• Why does the OCR result for that simple image depend on the resolution at all?