Devanagari handwritten digits recognition using weighted neighborhood self-organizing map

In the conventional SOM all the cells in the neighborhood of the winning neuron are updated by giving the same treatment to each of them. However, the proposed weighted neighborhood SOM (WNSOM) algorithm updates these cells by varying factor, which is a function of the distance of the neighboring neuron from the winning neuron and the current neighborhood radius. Both linear and exponential functions of these parameters are tried. The proposed procedure using these functions offered better results than conventional SOM. These results are also compared with Type-I Learning Vector Quantization (LVQ-1) and are found to be better than those obtained after fine-tuning, which requires thousands of iterations applied to the initial map created using the conventional SOM.