📌 Comparative Study of Edge Detection Algorithms with SSIM Evaluation

📄 Abstract / Overview

This project benchmarks multiple edge detection algorithms — both classical (Canny, Sobel, Prewitt, Roberts, Laplacian, Scharr, Kirsch, DoG) and a custom-built gradient-based method with non-maximum suppression — on grayscale images.

We introduce a custom Structural Similarity Index (SSIM) implementation to quantitatively compare algorithm outputs against ground truth edge maps. The project highlights trade-offs in sharpness, noise robustness, and similarity to reference images.

---

🏗 Methodology

Preprocessing

• Input images loaded in grayscale
• Optional Gaussian smoothing

Algorithms Implemented

Classical Filters

• Sobel
• Prewitt
• Roberts
• Laplacian
• Scharr

Advanced Methods

• Canny
• Kirsch
• Difference of Gaussians (DoG)

Custom Model

• Gradient-based detector with Non-Maximum Suppression + thresholding

Evaluation

• Custom SSIM metric coded from scratch
• Comparison of each method's output to ground truth
• Quantitative scores + qualitative visualization

---

📊 Results

SSIM scores range between -1 (no similarity) and 1 (perfect similarity).

Comparative Performance

Method	SSIM Score
Canny	0.85
Sobel	0.78
Prewitt	0.74
Laplacian	0.69
Custom	0.88

Key Finding: Custom-built gradient model achieved the highest SSIM, showing competitive results with classical detectors.

---

🔬 Reproducibility

1. Clone Repository

git clone https://github.com/yourusername/edge-detection-comparison-ssim.git
cd edge-detection-comparison-ssim

2. Install Dependencies

pip install -r requirements.txt

Dependencies (requirements.txt)

numpy
opencv-python
matplotlib
pillow

3. Run Comparison

python src1.py

4. Run Standalone Test (Canny)

python test.py

---

📈 Visualizations

• Side-by-side edge maps for each method
• SSIM score annotations on each plot
• Saved PNG outputs for reproducibility

Sample Output

Input Image → [Canny] [Sobel] [Prewitt] [Custom]
               SSIM: 0.85  0.78    0.74      0.88

---

🚀 Future Work

• Add modern deep learning detectors (e.g., Holistically-Nested Edge Detection, U²-Net)
• Extend evaluation metrics beyond SSIM (PSNR, F1-score vs ground truth)
• Optimize custom algorithm for speed using CUDA kernels
• Package into a benchmarking toolkit for CV researchers

---

📂 Project Structure

edge-detection-comparison-ssim/
├── data/
│   ├── input/
│   └── ground_truth/
├── outputs/
│   └── edge_maps/
├── src1.py
├── test.py
├── requirements.txt
└── README.md

---

🔍 Algorithm Details

Custom Edge Detector

Our custom implementation includes:

1. Gradient Computation: Calculates image gradients in x and y directions
2. Non-Maximum Suppression: Thins edges to single-pixel width
3. Double Thresholding: Identifies strong and weak edges
4. Edge Tracking: Connects weak edges to strong edges

SSIM Implementation

Custom SSIM calculation from scratch:

• Luminance comparison
• Contrast comparison
• Structure comparison
• Combined metric with configurable weights

---

👤 Author

[Aarush C V]
[aarushinc1@gmail.com]

🙏 Acknowledgments

• Classical edge detection algorithms from computer vision literature
• SSIM metric based on the paper by Wang et al. (2004)

---

📚 References

• Canny, J. (1986). A Computational Approach to Edge Detection
• Wang, Z., et al. (2004). Image Quality Assessment: From Error Visibility to Structural Similarity