Artificial Intelligence, 2020 Fall, SJTU
by Prof. Jun Z.
For more details, please refer the report.
- Understand the provided codes of
OpenCV Sudoku Solver and OCRapproach. - Cooperate with fellow classmates to construct the
EI339-CNdataset of handwritten Chinese numbers一, 二,..., 九, 十. - Implement LeNet-5 and train on
MNIST + EI339-CN. - Implement a CSP solver for the Sudoku problem.
- Test the trained model and the solver upon given test images.
- Fully-annotated codes: both in function prototypes and between the lines.
- Carefully designed and wrapped functionalities: divide the board-extraction, Sudoku-board-formation, problem-solving processes into different packages and provide high-level APIs.
- Extensive experiment upon (1) influence of the selection of train and test datasets; (2) SudokuNet VS LeNet-5; (3) LeNet-5 activations (with analysis), training hyper-parameters.
- Error-tolerable CSP Sudoku solver.
- Construct real-world Sudoku problems set based on the published book
《全民数独》and conduct tests upon the images captured from 5 angles: bird-view, left, upper, right and lower. - Analysis and solid intermediate results.
- Some other implemented useful functionalities.
~
│ .gitignore
│ LICENSE
│ README.md
│
├─digit_classifiers // codes for digit classifiers
│ networks_models.py // network utilities (e.g. training)
│ networks_structures.py // network structures (LeNet-5 here)
│ __init__.py
│
├─extract_sudoku // [wrapper] step of extract Sudoku image
│ board_img_arr.py // image == extract ==> Sudoku board object
│ LeNet5__predict.py // API of LeNet-5 prediction
│ LeNet5__train.py // API of LeNet-5 training
│ opencv__predict.py // API of SudokuNet prediction
│ opencv__train.py // API of SudokuNet prediction
│ __init__.py
│
│─solve_sudoku // codes for Sudoku solver
│ sudoku_board.py // implementation of Sudoku board
│ sudoku_solver.py // Sudoku solver based on the board object
│ __init__.py
│
│
├─examples // example codes
│ extract_board.py // example codes to extract Sudoku boards
│ solve-sudoku.py // example codes to solve Sudoku problems
│ sudoku_puzzle.jpg // example image
│
├─run // final codes to execute
│ │ final_imgs.py // solve the given 5+5 final images
│ │ LeNet5_params.py // hyper-parameters test codes
│ │ LeNet5_structure.py // structure test codes
│ │ solve_by_LeNet5.py // LeNet-5 100-image test codes
│ │ solve_by_opencv.py // SudokuNet 100-image test codes
│ │
│ ├─final_out // results of the given 5+5 final images
│ ├─num_changed_plot // results of 100-image test
│ ├─params_plot // plots of hyper-parameters tests
│ ├─params_report // .json reports of hyper-parameters tests
│ ├─structure_plot // plots of network structure tests
│ └─structure_report // .json reports of structure tests
│
│
├─data
│ │ load_local_dataset.py // data loader codes
│ │
│ ├─EI339-CN dataset sjtu // EI339 dataset folder
│ │ ├─1, 2, ..., 10 // raw data of EI339
│ │ ├─mapping // intermediate folder: filenames mapping of EI339
│ │ └─processed // data file (.pt) of EI339
│ │
│ ├─MNIST // data file of MNIST
│ └─MNIST+EI339 // combined data file (.pt) of MNIST+EI339
│
├─imgs // test images folder & loader
│ │ load_imgs.py // image loader
│ │ overview.py // generate an overview of 100-image test images
│ │ sudoku_puzzle.jpg // test image
│ │ __init__.py
│ │
│ ├─JilinUnivPr // images captured from 《全民数独》, 吉林大学出版社
│ │ └─imgs // 100 images of 100-image tests
│ │
│ └─test1 // final 5+5 test Sudoku problems images
│
├─models // trained models
│
├─opencv_sudoku_solver // given OpenCV approach proposed online
│ │ solve_sudoku_puzzle.py // solved codes
│ │ sudoku_puzzle.jpg // test image
│ │ train_digit_classifier.py // train codes
│ │ __init__.py
│ │
│ ├─output // output models
│ │
│ └─pyimagesearch // SudokuNet & captured board extractor
│ │ __init__.py
│ │
│ ├─models // SudokuNet structure
│ │ sudokunet.py
│ │ __init__.py
│ │
│ └─sudoku // captured board extractor
│ multi_img_view.py // added functionality: combine & show multiple images together
│ puzzle.py // board extractor: find puzzle & extract digit
│ __init__.py
│
└─report // report files