K-means Clustering Solution for Image Compression

The paper presents a detailed research study of the k-means clustering algorithm to be used for image compression tasks, where the RGB values of the colors are considered XYZ coordinates of 3D points which must be clusterized. A first step was to deeply understand algorithm's features and the most reasonable method was to create an original ‘from scratch’ implementation (GNU Octave) in order to deal with all its particular aspects and to master the details. The application was extensively tested and plans for further development were done. An alternate solution was to find an inspiring environment where the k-means method was implemented, in this way being studied the scikit-learn library. An original software was developed (Python) and after extensive tests, the conclusion was that the facilities of the scikit-learn solution were same with the ones already included in the development plan that followed the ‘from scratch’ implementation. Using a ‘trial-and-error’ approach, the scikit-learn based original application was used to find the most appropriate number of colors which keeps a reasonable accuracy of the initial image and also decreases the size of the compressed file. An additional idea to minimize the size of the output image is to use less than 8-bit representation for the few colors in the compressed file, e.g. 3-bit for 8 colors or less. To conclude, there is no general recipe to find an optimal solution for image compression using k-means. Therefore, original software instruments (GNU Octave, Python) are useful for an exploratory strategy to find, at least, a better suboptimal solution. Moreover, for practical reasons, they are reusable software libraries to be employed in follow-up research studies.