https://github.com/alsult/compvision_puzzle

A computer vision pipeline that reconstructs a shuffled puzzle using SIFT feature matching, RANSAC, and image preprocessing.
https://github.com/alsult/compvision_puzzle

computer-vision contour-detection feature-matching homography image-processing image-segmentation opencv peppa-pig puzzle-reconstruction python ransac-algorithm sift-features

Last synced: 12 days ago
JSON representation

A computer vision pipeline that reconstructs a shuffled puzzle using SIFT feature matching, RANSAC, and image preprocessing.

• Host: GitHub
• URL: https://github.com/alsult/compvision_puzzle
• Owner: AlSult
• License: mit
• Created: 2025-03-25T11:02:18.000Z (about 2 months ago)
• Default Branch: main
• Last Pushed: 2025-04-23T08:05:46.000Z (26 days ago)
• Last Synced: 2025-04-23T09:21:44.849Z (26 days ago)
• Topics: computer-vision, contour-detection, feature-matching, homography, image-processing, image-segmentation, opencv, peppa-pig, puzzle-reconstruction, python, ransac-algorithm, sift-features
• Language: Python
• Homepage:
• Size: 5.84 MB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# Peppa Puzzle Reconstructor

This project automatically reconstructs a shuffled image of a Peppa Pig puzzle using computer vision techniques. The algorithm identifies and extracts puzzle pieces, matches them with their correct positions in a reference image using feature matching and RANSAC, and overlays them to restore the original image.



  

  

  



## Preprocessing

Before reconstruction can take place, the shuffled puzzle image undergoes several preprocessing steps to isolate individual puzzle pieces:

1. **Grayscale Conversion**  

   The input image is converted to grayscale to simplify edge detection and reduce computational load.

2. **Edge Detection**  

   Canny edge detection is applied to highlight the contours of each puzzle piece.

3. **Morphological Operations**  

   Dilation and closing operations are used to strengthen the edges and close small gaps, making contours more distinct.

4. **Contour Detection**  

   External contours are identified to extract bounding boxes around each puzzle piece.

5. **Filtering by Area**  

   Small contours (noise or irrelevant fragments) are filtered out using a minimum area threshold to retain only meaningful puzzle segments.

The result of this preprocessing step is a set of cropped images — each representing a potential puzzle piece — which are then passed on for feature matching and reconstruction.

### Visual Output of Preprocessing



  

  

  



  

## How It Works


1. **Piece Extraction**  

   Detects individual puzzle pieces from a shuffled image using edge detection and contour analysis via OpenCV.



  

  

  



3. **Feature Matching with RANSAC**  

   Each puzzle piece is matched to a reference image using SIFT keypoints and descriptors. RANSAC is applied to estimate a homography matrix for accurate alignment.



  

  



4. **Transparent Background**  

   White areas in the puzzle pieces are made transparent to blend seamlessly during reconstruction.



  



5. **Puzzle Reconstruction**  

   Each aligned piece is overlayed onto a transparent canvas, rebuilding the complete puzzle.



  



## Results

The reconstruction works even when the pieces are rotated or slightly misaligned, thanks to the robustness of the SIFT + RANSAC approach. While it performs well on clean images with distinct features, accuracy may decrease with blurry or low-contrast images.   

## Limitations

While the reconstruction pipeline is generally accurate, there are cases where two connected puzzle pieces are mistakenly treated as a single object during contour detection. This typically happens when:

- The gap between pieces is too small or visually subtle

- Morphological operations (dilation, closing) merge adjacent edges

- The white background isn't fully separated from piece boundaries

As a result, a few puzzle segments may be incorrectly aligned or misplaced in the final reconstruction.

Some extracted regions contained multiple puzzle pieces, as it's shown during the detection phase:

![Detected Pieces](extracted_pieces/detected.png)

Future improvements could include:

- Adaptive contour separation techniques

- Color-based segmentation to better split adjacent regions

- Manual validation for ambiguous cases

## Files Used

- `peppa/peppa.png`: The original complete puzzle image (reference).

- `peppa/pz.png`: The image containing randomly placed/shuffled puzzle pieces.

## Requirements

- Python 3

- OpenCV (with `cv2.SIFT_create()` support, e.g., via `opencv-contrib-python`)

- NumPy

Install dependencies:

```bash

pip install opencv-contrib-python numpy