Frequency sensitive self-organizing maps and its application in color quantization

A competitive learning algorithm named frequency sensitive self-organizing maps (FS-SOM) is proposed. It harmonically blends the neighbourhood adaptation of the well-known self-organizing maps (SOM) with the neuron dependent frequency sensitive learning model. The net effect is an improvement in adaptation, a well-ordered codebook and the alleviation or underutilization problem. A global butterfly jumping sequence is used to permute the input data and the dead neurons are re-initialized in the early phase of the training process. Extensive simulations have been performed to analyze and compare the learning behaviour and performance of FS-SOM against other vector quantization algorithms. The results show that FS-SOM is superior in reconstruction quality and topological ordering and its performance is highly robust against variations in network parameters.