Partially Hidden Markov Models (PHMM) are introduced. They differ from the ordinary HMM's in that both the transition probabilities of the hidden states and the output probabilities are conditioned on past observations. As an illustration they are applied to black and white image compression where the hidden variables may be interpreted as representing noncausal pixels.
|Journal||I E E E Transactions on Information Theory|
|Publication status||Published - 1996|