The Application of Multiple ANN/HMM Hybrid Models in Handwriting Recognition

摘要

In the issue of model recognition, big differences exist among the samples because of many factors such as the targets and environments in collection activities.The differences lead that it is very difficult for the single-parameter models to be trained to the appropriate values for building more accurate models.In view of the issue that the samples in sample sets have many differences, this paper comes up with the idea of multiple ANN/HMM hybrid models on the basis of ANN/HMM hybrid models. In this kind of models, each sub-model can assemble the samples that have little differences.And in the process of training and recognizing, the sub-models that can win the samples are the winners in the completions.In the single ANN/HMM hybrid model, the artificial neural network is adopted as the predictive network to calculating the output errors in each state of Hidden Markov Model.While in the multiple ANN/HMM hybrid models, the competitive mechanisms, which are introduced to each sub-model, will integrate all sub-models as a whole to build models for a mode.In this paper, the multiple ANN/HMM hybrid models are applied in handwritten symbols recognition.The same symbol may have many differences for no one has the same writing habit- the fact meets the objective requirement of building multiple ANN/HMM hybrid models.