Scanner_OP

1. Canny example 1

Taken from https://www.pyimagesearch.com/2014/09/01/build-kick-ass-mobile-document-scanner-just-5-minutes/
added median blur and tweaked few parameters.
Results are simply horrible.
Works only for 1 of the examples.
Transform library provided also yields the most horrible results I have ever seen.
Run it with python src/ex01/canny.py --image data/img_02.jpg

Taken from https://stackoverflow.com/a/18908023/8558040
Much better than 1st example but doesn't work on all.
Run it with python src/ex02/canny.py data/img_02.jpg
The updated code is in the jupyter notebook.
Improved the original code with median filter and dilation with unit kernel.
Tried to add code for selection the best 4 points if the contour has more and failed. Will try again with more robust geometry.
Will study bilateral filter, could be an improvement over gaussian filter in preserving edges.
I am postponing it and will come later after trying other examples.

Taken from https://stackoverflow.com/questions/32913157/how-to-get-magic-color-effect-like-cam-scanner-using-opencv
Comparing Global V/S Adaptive thresholding
Involves: Grey scaling, thresholding, denoising
Run it with python src/ex03/thresholding.py data/img_03.png

Taken from https://stackoverflow.com/questions/32913157/how-to-get-magic-color-effect-like-cam-scanner-using-opencv#62634900
https://answers.opencv.org/question/66125/object-detection-in-nonuniform-illumination/#79578
Was hoping for shadow and paper cringes removal.
Didn't get satisfactory results.
Tried
- Histogram equalization
- CLAHE
- Tophat addition and Blackhat removal
- Adaptive Threshold
- fastNlMeansDenoising
https://stackoverflow.com/questions/49997681/scanned-document-text-background-clarity-not-good-using-opencv-ios in file filter.cpp yields very promising results. There is an unimplemented part related to no text found in the image.
Tried Flat field correction - doesn't work so well.
Best version till now on non-white b.g. is to simply multiply the image by 1.5 (and subtract 40, somehow this lead to bit more noise but preserves information that is otherwise lost), basically increase contrast.

https://github.com/mauvilsa/imgtxtenh
The procedure for compiling is:
- cd build
- cmake -DCMAKE_BUILD_TYPE=Release ..
- make
Run with ./imgtxtenh img_path
All the important functions from the purpose of the project are defined in filters.hpp and documented in filters.cpp
Good parameters for magic filter usage will be alpha=1.4, beta=40

C++ implementation of ex02 with modifications learned during experiments.
The procedure for compiling is: g++ main.cpp corners.cpp -o main.o `pkg-config --cflags --libs opencv` -DDEBUG
Run with ./main.o [img_in_path] [img_out_path] [] is for optional arguments.
Implements the function find_best_corners which takes in img and returns the best rectangle the algorithm managed to find.
Implements four point transform for cropping and warping.
Implements order points.
Implements validate points.
Completed with documentation.

User manual from camscanner - http://s.intsig.net/r/manual/CamScanner_Android3.0_en-us.html
http://www.recogniform.net/eng/how-to-enhance-scanned-documents.html
Document Image Processing for Scanning and Printing by Ilia V. Safonov, Ilya V. Kurilin, Michael N. Rychagov, Ekaterina V. Tolstaya. Added book in Resources. (Who's going to read this. LOL)