Home > Howto > Scanning and recognition quality

Optimizing recognition results - Scanning and resolution

May 8th, 2012
The quality of recognition output is strongly influenced by image quality. Well illuminated images in good resolution produce few recognition errors - and vice versa. This page collects some tips that may help you to improve the quality of recognition results right from the beginning.

We will demonstrate some of the effects of different settings with a few lines from this test image (click to enlarge):

Running HindiOCR on this image produces the following text:

और भी ज्यादा मेहनत करने की जरूरत थी । इब्राहीम शरीफ
की कहानी ' खाली समय' (दीपावली विशेषांक) अधिक
अच्छी थी, परंतु उनकी यह कहानी (धूप का मैदान) भी
अच्छी है हालांकि 'खाली समय' से थोड़ा कम ।
Which factors influence the good recognition quality?

Image resolution

The test image is scanned with 300 dpi. The same image scanned with 100 dpi (click to enlarge) ...

... produces an output that has a lower accuracy rate.

Gradient at the book fold

Especially when scanning double pages, you frequently observe a fold-gradient, i.e. a dark area at the center area of the page, as seen in the following image (click to enlarge):

This gradient can influence recognition quality.

OCR for Hindi and Sanskrit offers a special function to suppress this effect of scanning. Using the command "Equalize illumination", the program is able to recognize the letters close to the book fold with higher accuracy.

Greyscale or black-white?

Try to use a greyscale image whenever possible! The Devanagari OCR programs rely on greyscale information to reconstruct fine features of images that may get lost when the image is scanned in black-white mode.

News

January 5th, 2016: Gujarati OCR (1.0.0.1) released

August 4th, 2015: Marathi OCR (1.0.0.4) released

All news and additional content