A computer vision project implementing two advanced image processing algorithms: GrabCut for automatic foreground segmentation and Poisson Image Editing for seamless image composition. This project demonstrates expertise in image processing, graph algorithms, and numerical optimization.
This project consists of two main components:
-
GrabCut Algorithm: An interactive foreground segmentation technique that uses Gaussian Mixture Models (GMM) and graph cuts to automatically extract objects from images with minimal user input (just a bounding box).
-
Poisson Blending: A sophisticated image compositing technique that seamlessly blends a source image into a target image by solving Poisson equations, preserving gradients and creating natural-looking results.
The GrabCut algorithm successfully segments foreground objects from complex backgrounds using only a bounding box as input.
| Original Image | Segmented Result |
|---|---|
 |
 |
| Book | Segmented Book |
 |
 |
| Flower | Segmented Flower |
 |
 |
| Teddy Bear | Segmented Teddy |
Poisson blending seamlessly composites segmented objects into new backgrounds while preserving natural lighting and texture.
| Blended Result | Background Image |
|---|---|
 |
 |
| Sheep in Mountain Landscape | Original Background |
The sheep has been seamlessly blended into the mountain landscape, maintaining natural lighting and texture consistency.
- Graph-based segmentation using min-cut/max-flow algorithms
- Gaussian Mixture Models (GMM) for foreground/background modeling
- Iterative refinement with convergence detection
- Energy minimization using N-links (neighborhood) and T-links (terminal) in a graph structure
- Achieves high accuracy (97%+ Jaccard index) on diverse test images
- Sparse matrix solving using SciPy for efficient computation
- Gradient preservation to maintain texture and lighting consistency
- Seamless boundary blending that eliminates visible seams
- Multi-channel support for RGB color images
- Python 3 - Core implementation language
- NumPy - Numerical computations and array operations
- OpenCV - Image I/O and basic image processing
- scikit-learn - Gaussian Mixture Models and K-Means clustering
- SciPy - Sparse matrix operations and linear algebra solvers
- igraph - Graph data structures and min-cut algorithms
├── src/
│ ├── grabcut.py # Main GrabCut algorithm implementation
│ ├── grabcut_utils.py # Helper functions for GMM and beta calculation
│ ├── graph_utils.py # Graph construction and min-cut computation
│ └── poisson_blending.py # Poisson image editing implementation
├── data/
│ ├── imgs/ # Source images
│ ├── bg/ # Background images for blending
│ ├── bboxes/ # Bounding box coordinates
│ ├── seg_GT/ # Ground truth segmentation masks
│ └── results/ # Generated results
│ ├── GrabCutResult/ # GrabCut segmentation results
│ └── PoissonBlend/ # Poisson blending results
└── tests/ # Test scripts and evaluation
- ✅ Automatic Segmentation: Extract objects with minimal user interaction
- ✅ High Accuracy: 97%+ Jaccard index on test dataset
- ✅ Seamless Blending: Natural-looking image composition
- ✅ Robust Implementation: Handles various image types and complexities
- ✅ Efficient Processing: Optimized algorithms for practical use
- ✅ Comprehensive Testing: Test suite covering multiple scenarios
- Initialization: User provides a bounding box around the object
- GMM Training: Initialize foreground and background Gaussian Mixture Models
- Graph Construction: Build a graph with pixels as nodes, connecting neighbors (N-links) and to source/sink (T-links)
- Min-Cut Computation: Solve for optimal segmentation using graph cuts
- Iterative Refinement: Update GMMs and repeat until convergence
- Mask Definition: Define the region of interest using a binary mask
- Gradient Computation: Calculate gradients from the source image
- Poisson Equation Setup: Construct sparse linear system Ax = b
- Sparse Solver: Solve for pixel values using sparse matrix techniques
- Boundary Conditions: Blend with target image at boundaries
python src/grabcut.py --input_name book --use_file_rect --evalpython src/poisson_blending.py --src_path data/imgs/sheep.jpg --mask_path data/seg_GT/sheep.bmp --tgt_path data/bg/grass_mountains.jpeg- Segmentation Accuracy: Average Jaccard index > 97% across test dataset
- Processing Speed: Optimized for practical use with convergence detection
- Memory Efficiency: Sparse matrix representation for large images