Quality and size assessment of quantized images using K-Means++ clustering

Davin Ongkadinata, Farica Perdana Putri

Abstract


In this paper, an amended K-Means algorithm called K-Means++ is implemented for color quantization. K-Means++ is an improvement to the K-Means algorithm in order to surmount the random selection of the initial centroids. The main advantage of K-Means++ is the centroids chosen are distributed over the data such that it reduces the sum of squared errors (SSE).  K-Means++ algorithm is used to analyze the color distribution of an image and create the color palette for transforming to a better quantized image compared to the standard K-Means algorithm. The tests were conducted on several popular true color images with different numbers of K value: 32, 64, 128, and 256. The results show that K-Means++ clustering algorithm yields higher PSNR values and lower file size compared to K-Means algorithm; 2.58% and 1.05%. It is envisaged that this clustering algorithm will benefit in many applications such as document clustering, market segmentation, image compression and image segmentation because it produces accurate and stable results.


Keywords


Color quantization; Image compression; K-Means++; Machine learning; True color image

Full Text:

PDF


DOI: https://doi.org/10.11591/eei.v9i3.1985

Refbacks

  • There are currently no refbacks.




Bulletin of EEI Stats