This paper presents a hybrid framework of feature extraction and hidden Markov modeling (HMM) for two-dimensional pattern recognition. Importantly, we explore a new discriminative training criterion to assure model compactness and discriminability. T