An Adaptive HMM Approach To On-Line Hand-drawn Shape Recognition

In this paper, an adaptive Hidden Markov Model (AHMM) approach for on-line hand-drawn shape recognition is presented. In our method, HMMs are chosen as the core recognizer due to its great ability to model stochastic time series. Many improvements are made to the traditional HMM recognizer in order to increase the flexibility of the recognition system, the resulting framework can not only adaptively learn from training data, but also can adapt system behavior to input hand-drawn shape. In order to fitting original hand-drawn graphics data to HMMs, effective pre-processing algorithms are designed to remove noise and rearrange input data points. Features for HMM training and recognition are carefully selected according to characteristics of HMMs, and corresponding feature extraction algorithms are introduced. For better performance in the circumstance of different hand-drawn shape input, an adaptive HMM (AHMM) structure is presented which combines single-band integral algorithm and an adaptive compression ratio control technique into one closed loop feedback recognition system, the AHMM can adaptively compress feature vectors according to geometry feature of input graphics, also can it adaptively adjust the feature compression ratio by feedback, this recognition structure achieve good performance in recognition. Taking the real-time recognition request into consideration, a bi-layer AHMM model (BL-AHMM) is introduced. BL-AHMM can reduce the recognition space by pre-categorization, which improves the speed and accuracy of recognition. Like the other grads descend algorithms, the problem of local optima is also existed in the training of HMM. In order to alleviate the influence to the recognition system, a hybrid GA-HMM training model is also presented, which introduce the genetic algorithm into the HMM training to search for global optima, the training performance is greatly improve by this model. Lots of experiments are performed in this paper, the result of these experiments shows that the BL-AHMM model trained by GA-HMM model is effective and has better performance over the traditional HMM recognition method.