Change KMeans Initialization method for Better Image Segmentation

[truongd3]

Currently, the KMeans clustering algorithm initializes centroids by selecting random points. This approach may lead to poor convergence, suboptimal cluster assignments, and inconsistent segmentation results.

Acceptance

• Ensures better spread of initial centroids
• Replace random centroid selection with "Perform KMeans clustering on a subset of pixels and use the result as the initial means"

Benefits

• More stable clustering results
• Faster convergence
• Better image segmentation quality