Tesseract: Tesseract is very poor on single line images

[Mohamed209]

i am running tesseract on some images to extract text in it , although images preprocessing seems optimal , tesseract cant detect any character in these images although i have tried different PSM and resizing options and i tried to finetune the model with such images type but also no results
check images below







does this issue related to inverting the image , because the original images were something like that


tesseract 4 and newer doesnot detect white text on black background , so i crop the black part invert it and send it to the engine , but no results

[stweil]

Well, at least some characters are recognized:

$ tesseract 64711738-69782d80-d4ba-11e9-80d2-23184f8c287b.jpg - --psm 7 -l tessdata_best/script/Latin
Warning: Invalid resolution 0 dpi. Using 70 instead.
w Tota e R

[Mohamed209]

@stweil i was thinking that the image is to easy , the application that i am working on needs all characters to be extracted correctly

[zdenop]

I am affraid tesseract works according rule: "Bad in - bad out":

tesseract cropped_total.png - --psm 6
Total Due: 60.

tesseract cropped_total2.png - --psm 6
Total Due: 297.50

[Mohamed209]

@zdenop
what tesseract version and langdata you used

[zdenop]

I used tesseract 5.0.0-alpha-329-gc5a50 and tessdata_best.
IMO preprocessing is more useful than training (especially if you are able to do text detection outside of tesseract is cases like you shown here)

[Mohamed209]

@zdenop did you preproceesd the image after cropping the borders ?

[zdenop]

First image worked without anything else just cropping.
Second image I need to dewarp (fix slope) first, so cropping fit to rectangle...
Tesseract has problem with borders: just search for "table" in issue tracker...

[Mohamed209]

@zdenop but it is not practical to just fit the text without any borders , although after many preprocessing some horizontal or vertical lines / blobs may residue in the image
do you see downgrading to versions that supports white text on black background may help

[zdenop]

Then tesseract is not right tool for you task. Tesseract needs straight text without border& graphics elements (+ 4.x version need black text on white background after thresholding) to provide good result. It is not bug, it is requirement.

[NguyenThaiHoc1]

@zdenop you can share which your code preprocess I would like to learn more from your code
Thank you for reading
Thai Hoc

[Mohamed209]

@NguyenThaiHoc1 i have made all preprocessing , he just manually cropped the image to remove black borders