【作者】 李杰；

【导师】 王家廞；

【作者基本信息】 清华大学 ， 计算机科学与技术， 2004， 博士

【摘要】 图像识别技术在生产生活中有着十分广泛的应用,复杂图像的识别问题一直是近年来模式识别和图像处理研究的热点。本文基于隐马尔可夫模型进行图像识别问题的研究。如何将已经在语音识别领域取得了广泛应用的隐马尔可夫模型成功引入到图像识别中来是近年来相关研究讨论的热点。本文认为,针对不同对象建立合适的伪二维隐马尔可夫模型是目前最可行和有效的方法。本文首先提出了一种适用于脱机手写体汉字识别的嵌套伪二维隐马尔可夫模型。嵌套伪二维结构是对二维结构的一种合理简化,而这种简化是基于对汉字结构的深入分析以及对邻点和邻域相关性的全面考察。这种模型在国标一级汉字字库上具有91.8%和95.2%的一选和十选识别率。本文对人脸识别问题进行了类似的研究,并提出了一种简化伪二维隐马尔可夫模型。这种模型在ORL人脸库上具有99.5%的识别率,而复杂度比已有的人脸隐马尔可夫模型小。针对隐马尔可夫模型在上述的具体应用中出现的问题,本文提出经典隐马尔可夫模型理论具有状态集固定的缺陷。这种缺陷影响了隐马尔可夫模型对随机信号建模的能力,并限制了基于隐马尔可夫模型的分类器的性能。因此,本文提出了一种全新的自适应隐马尔可夫模型,可表示为θ(N,A,B,π)。它具有可变的状态集,目的是为了自动匹配信号的“真正”隐含状态集,提取更多的结构信息。它采用了一种基于确定性退火全局优化技术的收缩训练算法,这种算法以最大后验概率为优化标准。实验结果表明该自适应模型具有理论和实践上的双重意义:更符合信号的内在结构,能提取更多的结构信息,可提高隐马尔可夫模型对随机信号建模的能力和在复杂图像识别应用中的性能。本文提出的嵌套伪二维隐马尔可夫模型经过自适应改造后可使脱机手写体汉字的一选和十选识别率分别提高到95.9%和99.0%。更多还原

【Abstract】 The technology of image recognition is widely used in lives and industries. The recognition of complex images is now still the focus of the research on pattern recognition and image processing. This paper focuses on the research of image recognition based on the hidden Markov model (HMM). Recent years, people have put great effort into the research on how to introduce HMM to image recognition successfully that has been widely applied in speech recognition. This paper considers it the most feasible and efficient way to build proper pseudo 2D HMM for distinct object. This paper first proposes a nested pseudo 2D HMM for off-line handwritten Chinese character recognition. The nested pseudo 2D topological structure is a reasonable simplification of a truly 2D one. This simplification is based on an in-depth study of the Chinese character structure and a full-scale overview of the relationship of adjacent points and areas. The proposed model has a recognition rate of 91.8% and a top-ten correct rate of 95.2% on GB first-level character database. This paper also illustrates a similar research on face recognition and a simplified pseudo 2D HMM is put forward. This model has a high recognition rate of 99.5% on ORL face database and a low computational complexity in comparison with other face HMMs. Based on the questions of HMM brought by the above two applications, this paper points out the limit of fixed state set in classic HMM theory. Therefore, this paper proposes a novel self-adaptive HMM, which can be denoted by θ(N,A,B,π). It has a variable state set, for the purpose of automatically matching the “true” hidden state set and extracting more structure information of the signal being modeled. It employs a shrink training algorithm base on the deterministic annealing (DA) global optimization technique. This training algorithm adopts a maximum a posteriori (MAP) optimization criterion. Experimental results indicate that the self-adaptive model is <WP=6>useful in both the theory and application: it matches the inherent structure of signals better, can extract more structure information and can improve the performance of HMM in both stochastic signal modeling and image recognition applications. After the self-adaptive redesign, the proposed nested pseudo 2D HMM of the off-line handwritten Chinese character can increase the recognition rate to 95.9% and the top-ten correct rate to 99.0%.更多还原