CV Sudoku Solver

Photograph a Sudoku puzzle and get the solved grid as an image.

Input

Output

Model Performance

Grid Detection

• 99.5% mAP — detection accuracy across varying conditions
• 99.9% precision — minimal false positives
• 100% recall — never misses a grid

Digit Recognition

• 99% accuracy — correct digit on first prediction
• 100% top-5 — correct answer always in top 5 predictions

Robustness

• Printed and handwritten puzzles
• Varying lighting and exposure
• Skewed camera angles
• Any paper color or background

---

Models

Grid Detection Model

YOLOv8 Nano object detection trained on 500+ images from Roboflow.

Training Results

Training Curves	Precision-Recall Curve
Confusion Matrix	Validation Predictions

---

Digit Classification Model

YOLOv8 Nano classifier trained on 4,400+ cell images to recognize digits 1-9 and empty cells.

Training Results

Training Curves	Confusion Matrix
Validation Samples

---

How It Works

1. Detect — YOLOv8 locates the Sudoku grid in your photo
2. Warp — OpenCV corrects perspective to create a perfect square
3. Split — Grid is divided into 81 individual cells
4. Recognize — YOLOv8 classifier identifies each digit
5. Solve — Constraint propagation + backtracking finds the solution
6. Render — Solution is drawn onto a clean grid image

---

Solving Algorithm

Narrows down options dynamically before solving via depth-first search. Dynamic approach, narrow down options, then depth-first search

1. Constraint Propagation — Eliminates impossible values by checking rows, columns, and 3x3 blocks. Solves most puzzles without guessing.

2. Backtracking(DFS) — For harder puzzles, recursively tries possibilities and backtracks on invalid states. Guarantees a solution if one exists.

Solves any valid Sudoku in under 100ms.

---

Project Structure

backend/
├── sudoku_detector/     # Computer vision pipeline
│   ├── services/        # Detection, warping, recognition
│   ├── training/        # Model training scripts
│   └── models/          # Trained weights (.pt files)
├── sudoku_solver/       # Solving algorithms
└── sudoku_display/      # Solution rendering

---

Getting Started

# Clone the repo
git clone https://github.com/yourusername/cv-sudokusolver.git
cd cv-sudokusolver

# Install dependencies
pip install -r requirements.txt

# Run on an image
cd backend
python -c "import cv2; from main import solve_sudoku_image; solve_sudoku_image(cv2.imread('path/to/sudoku.jpg'))"

Output is saved to solved_sudoku.jpg.