Segmenting Coloured Images Using Bregman Divergence

Image segmentation is required for many applications of object recognition. It can be achieved through clustering in a partition based approach. K-means is a fast and efficient method for segmenting image into predefined number of regions. We propose how to improve its performance using Bregman divergences to measure similarity between two pixels. Results show that Bregman divergence is able to identify edges in more refined manner than Euclidean distance of classic k-means. The proposal is tested on several colored images to observe the visual separation produced and accuracy of segmentation is measured against human expert truth.