A mixture of local PCA learning algorithm for adaptive transform coding

Karhunen-Loeve transform (KLT) is the optimal linear transform for coding images under the assumption of stationarity. For images composed of regions with widely varied local statistics, R.D. Dony and S. Haykin (1995) proposed a transform coding method called optimally integrated adaptive learning (OIAL), in which a number of localized KLTs are adapted to regions with roughly the same statistics. The new transform coding method is shown to be superior to the traditional KLT. However, the performance of OIAL depends on an estimate of the global principal components of the data, which is not only computationally expensive bat also impractical in some cases. Another problem of OIAL is that the mean vector in each region is not taken into account, which is required to define a local PCA. The authors propose an improvement over the OIAL which replaces the winner-take-all (WTA) based clustering with an optimal soft-competition learning algorithm called “neural gas”. The mean vector in each region is also incorporated. Experiments show a better performance than OIAL