A self organizing map based pixel classification
Abstract
Two methods for pixel classification are presented in this article. These methods are very helpful for image compression, image restoration and image binarization. Both of these methods are based on Self Organizing Maps (SOM) . We exploit the ability of SOMs to preserve the topology of the input space ( the image) to classify the pixels based on the labeling of the SOM neurons. The first method (SOM_K) uses the Kmeans classifier to labelize the neurons of the SOM. A classification of the pixels is then performed via the labeled map. The second approach (SOM_K_MLP)goes further by adding a multi-layer perceptron which is trained on the SOM labeled neuron before classifying the pixels of the original image. The application of this last method on grey level images shows better results of binarization than classical methods