Sudoku Solver with Computer Vision & Deep Learning

Automated Sudoku puzzle solver using computer vision for grid extraction and deep learning for digit recognition.

Pipeline in Action

Original Image	Extracted Grid	Solved Sudoku
Raw input image	Detected & warped grid	Final solution overlaid

Features

• Image Processing: Automatic Sudoku grid detection and cell extraction
• Digit Recognition: CNN and ResNet152 models for handwritten digit classification
• Sudoku Solving: Backtracking algorithm for puzzle solving
• Transfer Learning: Pre-training on MNIST with fine-tuning on Sudoku datasets

Project Structure

src/
├── common/          # Utility functions (config, pickle)
├── core/            # Training logic
├── data/            # Dataset loaders (MNIST, Sudoku)
├── evaluate/        # Model evaluation scripts
├── model/           # Model architectures (ConvNet, ResNet152)
├── preprocess/      # Image processing pipeline
└── scripts/         # Main executable scripts
    ├── train.py     # Universal training script
    └── pipeline.py  # End-to-end Sudoku solver

Quick Start

Training a Model

Train ConvNet on MNIST:

python src/scripts/train.py --model convnet --dataset mnist --epochs 10

Train ConvNet on Sudoku:

python src/scripts/train.py --model convnet --dataset sudoku --epochs 50

Fine-tune MNIST model on Sudoku:

python src/scripts/train.py --model convnet --dataset sudoku --epochs 40 \
  --finetune models/10epochs_convnet_mnist.pkl --lr 0.0005

See TRAINING.md for more training examples.

Solving a Sudoku Puzzle

python src/scripts/pipeline.py

Or use in Python:

from src.scripts.pipeline import main_pipeline

solution = main_pipeline(
    image_path="path/to/sudoku.jpg",
    model_path="models/50epochs_convnet_sudoku.pkl",
    save_images=True,
    show_images=True
)

Evaluating Models

Evaluate all models in the models directory:

python src/evaluate/evaluate_all_models.py

Models

ConvNet

• Simple CNN architecture for 28x28 digit images
• 2 conv layers + 2 FC layers
• Fast training and inference

ResNet152

• Pre-trained on ImageNet with custom classification head
• Fine-tuned for digit recognition
• Higher accuracy but slower

Dependencies

• PyTorch
• OpenCV (cv2)
• NumPy
• Pandas
• scikit-learn
• PyYAML

Install with:

poetry install

Results

Models are automatically evaluated and saved in results/ directory with timestamps.

Pipeline outputs (original image, extracted grid, solution) are saved in results/pipeline_outputs/.