The application of agglomerative clustering in image classification systems

Agglomerative clustering is proposed as an unsupervised training method. The algorithm is controlled either by giving the number of clusters or by specifying some threshold value. In the latter case, the algorithm uses the adaptive threshold technique to achieve its natural clusterings. Similar to merging regions in image segmentation (M.D. Levine et al.. 1981), this method grows the clusters by attempting to merge as many logically adjacent pixels as possible, provided that the difference between each feature is less than some adaptive threshold value. In this study the algorithm was implemented by using both techniques. The classification results of the agglomerative method is compared with those of K-means and ISODATA training algorithms