K-Means Image Segmentation with RGB, Position, and Texture Features

Overview

A1. Implementing K-Means Clustering

• Segmentation was performed using K = 5 and K = 10 clusters.
• The stopping criterion was based on convergence of assignments or a fallback of 200 iterations (never reached).
• Images segmented:
  • sunset.tiff
  • tiger-1.tiff
  • tiger-2.tiff
• Segmentations are visually distinct, with higher K values capturing finer details.

A2. 5-Dimensional Feature Vector (R, G, B, λX, λY)

• Segmentation was repeated using a 5D feature vector combining color and pixel location.
• Weighting parameter λ was tested with values 1 and 10 for K = 10.
• Higher λ increases the influence of pixel position:
  • λ = 1: color dominates
  • λ = 10: spatial proximity dominates
• Observed that increasing λ causes more spatially localized segmentations.

A3. Using Texture Features

• Texture features computed using root mean squared values over a 10×10 window.
• Feature combinations tested:
  1. Texture only
  2. Texture + RGB
  3. Texture + RGB + position (λ = 1)
• All feature vectors scaled to [0, 255] for consistency.

Observations:

• Texture only: Captures some spatial structure but lacks sharp color segmentation.
• Texture + RGB: Adds subtle grouping but can degrade clean RGB-based segmentation.
• Texture + RGB + position: Adds spatial smoothing, but in these images, it reduces quality.

Additional Notes:

• Visual artifacts appeared near borders, possibly due to edge effects from convolution used to compute local texture.
• Artifacts changed with window size, suggesting that convolution padding or boundary handling in MATLAB might be responsible.

---

Conclusion

This assignment demonstrates that while K-means clustering is sensitive to feature representation, RGB values alone often produce the cleanest segmentations in color-dominated images. The inclusion of position and texture features introduces spatial and pattern sensitivity, which may be more useful in scenarios with low color contrast.