search

charlesw / tesseract

A .Net wrapper for tesseract-ocr
Apache License 2.0
2.27k stars 742 forks source link

Image format (specifics) for better (accurate) recognition? #444

Open ledzepff opened 6 years ago

ledzepff commented 6 years ago

Having A6 format ID, 150 dpi, B&W image and no skew, getting about 75% recognition accuracy.. Is there any way to increase it by reprocessing image as a preparation?

tdhintz commented 6 years ago

I find the details regarding how to process matter a lot. For example, there is a huge difference degradation when passing a Bitmap instead of having Leptonica load files directly to a Pix object. I don’t know why. Other areas where significant difference occur include your choice of segmentation mode and version (v4 LSTM ‘fast’ is often better than the old v3.x).

ledzepff commented 6 years ago

thank you for information. I read that v4 is a very slow in comparison to v3.x. BTW, passing a clear text within white background gives much more accuracy (that is the obvious I guess) and image pre-processing seems may be good (and painful and the same time) to have a more accurate recognition.

tdhintz commented 6 years ago

v4 LSTM is faster according to my benchmarks. Have you measured yourself?

ledzepff commented 6 years ago

I'm not, unfortunately, this is just I heard and it would be better for me to have my own tests.