sdjf wrote: ↑
Thu Feb 05, 2015 4:31 pm
THe problem I see is that OCR software can be pretty finnicky, and not do a very good job unless the quality of the image is very, very good, and perfectly level. I am not sure how good the OCR software is that is available for the Pi, but I think that is a very important factor.
chad546 wrote: ↑
Tue Aug 22, 2017 3:01 am
the problem in OCR api's is that it still read the words on the paper even though the blind person did not place the paper properly.
This might be the issue that can be usefully addressed - guiding the blind user to optimal positioning of the camera relative to the text using haptic (feel) or other non-visual feedback to enable effective OCR.
A Hough filter can find straight lines in images. OPenCV includes such filter functions. If the text is square-on and level the lines of text will be parallel and aligned with the pixel grid. If the view is off centre or rotated that can be measured.
A bounding box around the text should not overlap the edge of the image.
Isolating single words and attempting OCR on them as a fast rate may provide additional guidance before attempting the whole text.
A mechanism that moved or vibrated could guide the hand of the user to correct alignment. An inertial measurement unit (accelerometer) can measure shake and help to take a picture when the camera is steady.
The haptic devices might be vibration motors (as used in phones) or a moving weight (to feel the balance) or motors that apply force to the fingers touching the edge of the device or many other possibilities. Sound might also be useful.
Gloves may not be the best idea because they might interfere with touch and the device will probably not be in intensive use in the hand. For long reading sessions a table stand device is probably better. For occasional hand-held use a pendant on a lanyard might be easier.