Hi, my name is Rémi LUX. I end my final year of Computer Vision Degree in Grenoble INP - Phelma, and I'm now engineer at Niryo.
Welcome to my project
My algorithm is able to detect and to directly solve sudokus in VR
For a single image process, my algorithm is using working as following :
Firstly, the algorithm have to find where the grids are !
The pipeline is :
- Preprocess image to enhance high frequencies
- Find lines thanks to Hough Space
- Analyse lines to know which can correspond to a grid
Once grids are extracted, for each grid we do :
- Detect shape which can be digits
- Use my CNN do identify digits
- Create a numeric grid in a table & solve it
- Create a virtual image to fill the initial image
- Add the 2 images together to create the final result
When user chose to use video as input, the algorithm will act quite differently :
- For grids extraction, we use position of last grids to deduce if 2 grids in 2 pictures are actually the same
- For grids solving, we wait to have meet the same extracted numeric grid twice, to be confident on digits extraction
- It do not resolve a grid if it a grid is detected as same as an already solved grid
The best way to use my algorithm is to firstly clone the git
repo where you want to.
Then, open a terminal & go in the repository
cd /path/to/the/folder/
You can see the options by typing
python3 sudoku_solver_main.py -h
You will need some libraries before running it
- Numpy
- OpenCV
- Keras
- Tensorflow
To Install : pip install <library_name>
Take your own picture/video of your process target and use the command
python3 sudoku_solver_main.py -i path/to/your/file
You also want to add -s to save the output result
In both cases, the result will be display on your monitor during processing
If you want to use your webcam, you can chose the profile n°2
by adding -p 2 to your command. Nevertheless, the algorithm
is too long to give a real-time pleasant experience (mainly cause
of the Hough Transform)
The user can use his own Keras model to identify digits.
For that use the parameter -mp path/to/your/model
| Version | Source | Algorithm Output |
|---|---|---|
| 1.5 | Video |  |
| 1.3 |  |
 |
| 1.2 |  |
 |
| 1.1 |  |
 |
| 1.0 |  |
 |
I trained mine personally with on a dataset of 10k numeric digits with data augmentation. It gave me a precision of 99.5+ %
- More Robust (v1.2)
- Video Live Solving (v1.4)
- Better (v1.5)
- Stronger (v1.6)
- Faster (v1.7)
- Improve Training(v1.7)
-
Multi-threading Processing
- v1.7 : Multiple performance improvement | 17/12/19
- New training method
- Improve Reconstruction
- v1.6 : Multiple performance improvement | 19/05/19
- Optimize method for solving
- Optimize method for grid detection
- Improve Robustness It is now able do deal with Real-Time
- v1.5 : Stabilizing video resolution | 12/05/19
- Multiple checking if the grid is well detect
- Jump solving step if seems to be already solved !
- v1.4 : Video Handling | 08/05/19
- v1.3 : More flexible / New training, better CNN ! | 05/05/19
- v1.2 : Multiple grids baby ! | 02/05/19
- v1.1 : Use probabilistic Hough & detect grid better | 01/054/19
- v1.0 : First version | 28/04/19