Hidden control neural network and HMM hybrid approach for on-line cursive handwriting recognition

Abstract

The paper focuses on a hidden control neural network (HCNN) based ANN/HMM hybrid approach which handles simultaneously both the global pattern class variation and the local signal primitive variation. HMM is used at the pattern class level to organise different primitives in various orders. One HCNN is applied to model signal primitives in each HMM state as the emission probability estimator. The control signal of HCNN copes with the primitive variation absorption task. The proposed method was applied to the on-line cursive handwriting recognition problem and compared with our previous similar systems on the UNIPEN handwriting database.