Supervised color image segmentation, using LVQ networks and K-means. Application: cellular image

Abstract

This paper proposes a new method for supervised color image classification by the
Kohonen map, based on LVQ algorithms. The sample of observations, constituted by image pixels with 3 color components in the color space, is at first projected into a Kohonen map. This map is represented in the 3-dimensional space, from the weight vectors resulting of the learning process . Image classification by kohonen is a low-level image processing task that aims at partitioning an image into homogeneous regions. How region homogeneity is defined depends on the application. In this paper color image quantisation by clustering is discussed. A clustering scheme, based on learning quantisation vector (LVQ), is constructed and compared to the K-means clustering algorithm. It is demonstrated that both perform equally well. However, the former performs better than the latter with respect to the known number of although class. Both depend on their initial conditions and may end up in local optima. Based on these findings, an LVQ scheme is constructed which is completely independent of initial conditions; this approach is a hybrid structure between competitive learning and splitting of the color space. For comparison, a K-means approach is applied; it is known to produce global optimal results, but with high computational load. The clustering scheme is shown to obtain near-global optimal results with low computational load

Keywords: color image, kohonen, LVQ, classification, K-means