Image compression using a stochastic competitive learning algorithm (SCoLA)

Abstract

We introduce a new stochastic competitive learning algorithm (SCoLA) and apply it to vector quantization for image compression. In competitive learning, the training process involves presenting, simultaneously, an input vector to each of the competing neurons, which then compare the input vector to their own weight vectors and one of them is declared the winner based on some deterministic distortion measure. Here a stochastic criterion is used for selecting the winning neuron, whose weights are then updated to become more like the input vector. The performance of the new algorithm is compared to that of frequency-sensitive competitive learning (FSCL); it was found that SCoLA achieves higher peak signal-to-noise ratios (PSNR) than FSCL.