基于GA-AHMM的联机手绘图形识别技术

Online hand-drawn graphic symbol recognition based on gybrid model of GA and adaptive HMM

提出了一种遗传算法(geneticalgorithm,GA)和自适应隐马尔科夫模型(hiddenMarkwmodel,AHMM)混合的联机手绘图形识别方法。由于隐马尔科夫模型(HMM)的训练本质上是一种梯度下降的优化方法,算法易陷入局部最优,影响了其应用。为此,采用GA训练HMM模型参数,并给出了GA和HMM的两种混合训练方式:前端GA HMM模型和内嵌式GA HMM模型,GA算法能随机地调整HMM模型训练的初始值,使HMM跳出局部最优,较好地克服了HMM训练容易陷入局部最优的问题。另外,采用带有反馈环节的闭环AHMM代替传统的开环前向HMM模型对手绘图形识别,改善了HMM的自适应能力,显著提高了对图形的识别率和识别速度。试验结果证明了方法的有效性。

A new hybrid model of genetic algorithm (GA) and adaptive hidden Markov model (AHMM) for on-line hand-drawn symbol recognition is presented. Because the training of hidden Markov model (HMM) is an optimization method using gradient descent, it could convergent to local optima and limit its applications. GA is adopted to train the parameters of HMM. Two hybrid training methods, GA forward-based GA-HMM model and GA embed-based GA-HMM model, are proposed. GA can be used to adjust the initial training parameters of HMM randomly. It can avoid falling of the training of HMM into local optima effectively. Meanwhile, the adaptive ability of HMM can be improved using closed-loop AHMM with feedbacd part instead of traditional opened-loop forward HMM to recognize hand-drawn symbols.GA-AHMM structure can significantly improve the speed and accuracy of HMM to recognize hand-drawn symbol. Several experiments show that GA-AHMM model is effective and has better performance over the traditional HMM recognition method.