| Author |
 Topic  |
|
|
TimHayes
5 Posts |
 Posted - Apr 20 2020 : 10:53:05
|
Hello,
I have recently inistalled the latest versions of ImageEN and IEVision.
I am unable to replicate the "fatser performance" in IEVision on any full page text scans.
I attach a zip with the OCR demo project which I have compiled and run under Win64. I am using the "mixed" eng.traineddata.
The legacy scan for the test page I supply "Test_Page_19.jpg" runs in about 2 seconds, the LSTM + Legacy in about 8-9 seconds, and the LSTM only in 7-8 seconds.
Plainly the legacy is a poorer quality and difficult to use on a production environment.
I can get a similar page scanned at Google with a turnround of about 2 seconds, which is a very high quality but at a cost. I have trialed other systems available online who also use Tesseract, and they provide very good quality in around 2 seconds.
Can anyone please point out where I am going wrong?
Many thanks.
Tim Hayes
attach/TimHayes/2020420105210_OCR_Test.zip
3105.68 KB |
|
|
xequte
38198 Posts |
Posted - Apr 21 2020 : 17:50:43
|
Hi Tim
The performance you are seeing with IEVision is Tesseract as compiled by Visual Studio. We're not doing anything that would slow that down.
That said, we did work through all the flags that Tesseract offers, and there is one that significantly reduces processing time by using all the available cores (reduces the processing time by about 55% on my 8-core system, or around 30% on my partner's 4-core system). We don't know if this introduces any compatibility issues, but you can email us to test it.
I don't know that we can compare our results with Google, who will have computers that us mere mortals can only dream of. They may also be doing some optimization to the images before processing.
Other things that can help:
- Don't use JPEG images as the compression artifacts slow down processing. Gray-scale or B+W tiffs are best
- You could try optimizing the image before processing to see if that reduces overall processing time (expanding color range, lowering resolution, etc). There is no magic bullet there, but if you are regular working with images from the same source, you might find an optimal image type
- Naturally, don't test your speeds under Delphi (which slows OCR by 50%)
- 64bit is faster than 32bit (by about 10%)
Nigel
Xequte Software
www.imageen.com
|
 |
|
| |
Topic |
|
|
|
Search
IEVision Performance Issue?
