Computer Vision: Thresholding and Segmentation

Authors

• Ibrahim Fateen
• Bassant Zaki
• Mohamed Hisham
• Yasmine Mahmoud

Overview

This task explores several classical and modern image segmentation techniques. The main objective is to separate regions of interest in grayscale and color images using thresholding and clustering-based methods.

Covered techniques include:

1. Otsu Thresholding (Global & Local)
2. Optimal Thresholding (Global & Local)
3. Spectral Thresholding (Binary & Multi-level)
4. K-Means Segmentation
5. Region Growing Segmentation
6. Mean-Shift Segmentation
7. Agglomerative Segmentation

1. Otsu Thresholding

Otsu's method finds a threshold value that minimizes intra-class variance and maximizes inter-class variance.

• Applies to histogram-based analysis.
• Efficient for bi-modal distributions.
• Can be extended to local thresholding using windowed regions.

2. Optimal Thresholding

An iterative inter-means method that updates the threshold based on mean values of pixel groups until convergence.

• Works on pixel values, not histograms.
• Applied globally or locally using small windows.
• Tolerant to intensity distributions not suited to Otsu.

3. Spectral Thresholding

Spectral thresholding segments images based on intensity using between-class variance analysis.

• Binary Mode: Separates pixels into foreground/background.
• Multi-Level Mode: Uses two thresholds for three-class segmentation.
• Fallback to binary if three-class separation is ineffective.

Challenges include high computational cost and increased histogram sensitivity for multi-class segmentation.

4. K-Means Segmentation

K-Means clusters pixels based on color similarity.

• Input space: RGB or LUV (optional for better perceptual clustering)
• Each pixel is a feature vector
• Random centroids, distance-based assignment, and iterative update
• Output: clustered image by dominant colors

LUV space sometimes oversegments or missegments due to uniformity sensitivity.

5. Region Growing Segmentation

An interactive method that starts from user-defined seed points and expands regions based on intensity similarity.

• Supports multiple seeds
• Uses 8-connected neighborhood
• Threshold-based inclusion and queue for processing
• GUI allows seed selection and visual feedback

Challenges include finding optimal similarity threshold and visualizing binary masks.

6. Mean-Shift Segmentation

A density-based clustering algorithm that shifts feature vectors toward the mode.

• No need to preset number of clusters
• Feature space: pixel values (+ optional spatial info)
• Bandwidth controls convergence and granularity
• Low bandwidth → many fine clusters, High bandwidth → fewer coarse clusters

Performance is highly dependent on bandwidth selection.

7. Agglomerative Segmentation

A bottom-up hierarchical clustering method where each pixel starts as a cluster.

• Distance-based merging of clusters using a priority queue (heap)
• Two modes: Vanilla (pixel-based) and Superpixel-initialized (faster but less accurate)
• Builds a binary tree structure representing cluster hierarchy

Vanilla mode yields better results but is computationally expensive, requiring image downscaling for feasibility.

Dependencies

• Python 3.8+
• NumPy
• OpenCV
• Scikit-learn
• Matplotlib
• PyQt (for region growing GUI)

Install dependencies with:

pip install -r requirements.txt

Conclusion

This task presents a comprehensive comparison of segmentation methods. From statistical thresholding to clustering and region-based approaches, each method offers distinct strengths depending on image content and computational constraints. Combined, they form a powerful toolkit for various segmentation applications.