Android Tesseract (Tess-two) OCR Poor results with Seven Segment

Question

I am working on an app where you take a picture of a device and it stores the information. However, due to the device using seven segment display / LCD i am struggling to get good results.

I have used Firebase OCR, Google Vision ML Kit, And Tesseract (Tess-two).

Each of them are providing poor results, especially Tesseract considering i have trained and fine tuned it to recognise 3 types of seven segment fonts to a 0.09 error rate.

I have also tried using other peoples traineddata, such as letsgodigital

I have converted the image to greyscale and then binarized it, i have also whitelisted only numbers and still can't detect the text properly. What else can i do to improve the results?

Results below are both tesseract.

Answer 1

Tesseract is good at recognizing documents. The example you gave contains pretty irrelevant information such as the machine and disturbing logo.

Doing the text localization to separate the seven-segment content out would be better a way to preprocess it.