Past accuracy was good, but recently accuracy is low

I used ocr space engine 3 for my university project and its
accuracy was good, but recently its accuracy quite low. Please fix that.

Hello, we are not aware of any change that would make the OCR engine 3 recognition worse.
→ Can you please post an example image that we can use for testing? Ideally it would be an image that gave you good results in the past, but not now? We will look into it ASAP!

i also faced the same issue.

the word “fabbricato“ was accurately extracted in the past, but now it recognize as “fabblicato“. Thank you.

****** Result for Image/Page 1 ******
INTELLIGENT LABEL SOLUTIONS

bikini/bikini

made in Sri Lanka/fabriqué au Sri Lanka/hecho en Sri Lanka/fabblicato in Sri Lanka/斯里兰卡制造 body/gusset/corps/gousset/cuerpo/inserto/ corpo/tassello/面料/内裆 57% cotton/coton/algodón/cotone/棉
38% modal/莫代尔 5% elastane/élasthanne/elastano/elastan/氨纶 exclusive of decoration/sauf décoration/no incluye la decoración/esclusa la decorazione/装饰除外

RN54867 CA67359 ID36015502 415968 12 25
S/P/CH/170/68A

1 Like

@Sanjuna_Nimandith Do you also have a test image for us?

@Akila_Yesan Thanks for the test image. I confirmed the issue with it!

Next step: We will use your image to debug the issue further. If you have another 1-2 test images, that would be helpful, too.

This is the only image where I encountered the issue. I’ve tested this image more than 1,000 times in the past, and it worked perfectly every time. The issue only started occurring from yesterday onwards.

We are still debugging this. This change is probably related to an ocr code update, at least the time matches our update time of the free servers.

If strict spelling is your priority, I can recommend to reduce the size of the image. Very large fonts can sometimes cause problems. If you need precise formatting and accent marks, then keep the current size.

This image does not have the spelling mistake (size reduced by 50%, → also much faster).

If anyone else found reduced ocr accuracy for OCR engine 3 since last week, please let us know. Our own tests look fine. So each of your samples gets included in our OCR quality test suite. This way we can make sure that future updates show no regression on your specific images :slight_smile: