CPSC 473 - Data Mining
This project uses cluster mining techniques to convert RGB and RGBA images to index color images, reducing the file size. either the K-Means or Mean Shift algorithm can be used to produce a compressed, paletted copy of their original image.
This method begins by distributing k "mean-points" through the color space in a way that spreads them out among the colors of the image. Then it iterates through the following process:
- Assign each color to its nearest mean-point
- For each mean-point, take the weighted average of all the colors assigned to it
- Move each mean-point to its average
- Repeat until no mean-point moves more than some predefined amount
| Original 24-bit RGB (10,891 kb) | K-means 4-bit Indexed (721 kb) |
|---|---|
 |
 |
By taking the averages of all points in the space, K-means will tend to do well representing the dominant colors of an image.
This method starts with centroids uniformly distributed in the color space, then it iterates through the following process:
- Find all the colors within a fixed distance of the centroid
- Move the centroid to the average
- Repeat until the centroid no longer moves
- Repeat for each centroid
- Prune the centroids that have converged to the same color
| Original 24-bit RGB (1032 kb) | Mean-Shift 8-bit Indexed (99 colors, 182 kb) |
|---|---|
 |
 |
For images with many distinct color clusters, mean shift is able to quickly determine how many colors are necessary, and what they should be.