MVP 03 - detect skewed lines

kolhizin / bujo

Proof of concepts and sharedlib for bujo app

0 stars 0 forks source link

MVP 03 - detect skewed lines #34

Closed kolhizin closed 5 years ago

kolhizin commented 5 years ago

Create tool to identify orientation of text and probability of any text at all.

DoD

python module with functions, that are used to identify region of interest
python notebook with sample
wiki page with description of methods

kolhizin commented 5 years ago

Created MVP code in module. Ran sample test suite. 0: very good, though detected 1 extra and missed short title (may be fixed by minvalue) 1: worst result, though 90% is correct; missed 1 title, 1 paragraph misrepresented as if angle is wrong (can be solved by regularization) 2: very good, detected 1 extra 3: good, angle detection fires too strongly, though 4: good, angle too weak, thus missed due to incorrect curve 5: very good 6: trained on this one, good 7: very good

Overall results -- good. Can be tuned by parameters

kolhizin commented 5 years ago

Timing:

Angle and line-delta detection -- 0.25s
Variance filter -- 0.05s
Region selection -- 0.1s
Support curves detection -- 1.05s

Longest part in 4 is local radon calculation. If can be sped up it would improve overall speed.

kolhizin commented 5 years ago

Change of image size:

decrease 2x -- breaks algorithm
decrease slighlty -- quality deteriorates, speed improved marginally (0.8s vs 1.1s)
increase 2x -- improves quality, speed drops 3x (3.5s vs 1.1s)

kolhizin commented 5 years ago

Updated description in https://github.com/kolhizin/bujo_poc/wiki/PoC:-word-detection