【作者】 胡勇；

【导师】 高隽；

【作者基本信息】 合肥工业大学 ， 信号与信息处理， 2004， 硕士

【摘要】 多传感器信息融合技术是智能信息处理领域中的一个研究热点。从多传感器视角观察待识别模式能够完整准确的反映模式特征，消除观察信息的不确定性，因此，基于信息融合的模式识别方法已成为模式识别的发展趋势之一。目前信息融合在模式识别的应用主要在特征层和决策层展开。 本文围绕信息融合技术的理论基础和应用问题，从特征层融合和决策层融合的角度入手，对信息融合应用于模式识别进行了研究。 首先，提出一种特征层融合模式识别的方法，定义“特征融合系数”对多传感器视角观察模式所得的不同特征进行融合，通过对不同特征赋以不同的特征融合系数，将多特征进行融合，得到待识别模式的融合特征，从而实现特征层融合。 其次，在决策层融合的理论框架下，对D-S证据理论应用于模式识别进行了讨论，针对其在实际应用中各证据体的基本概率赋值难于获取的问题，以神经网络分类器的输出为基础获取分类不确定性的基本概率赋值，并以此作为分类、决策的依据。 再次，在决策层融合的理论框架下，针对模糊逻辑方法在实际应用中存在现有推理方法计算过于简单，容易丢失大量有用信息的问题，采用D-S证据理论进行模糊推理，并模拟生物大脑的“注意”功能对重要程度不同的规则以不同的对待。 最后，引入基于指纹方向场和手形特征的多模态生物特征识别实验对上面的决策层融合识别方法进行验证，为信息融合在模式识别方法中的实用化做出一点探索。更多还原

【Abstract】 Multisensor information fusion is a hotspot in the field of intelligent information processing. Because observing patterns to be recognized from multiple sensors can well and truly reflect the features of them, eliminate the uncertainty of the information. Pattern recognition methods based on information fusion have already been one of the development trends of pattern recognition field. At present, multisensor information fusion in pattern recognition is mainly applied in feature level and decision level.In this thesis, centering on the theoretical fundamental and application of multisensor information fusion technology, the applications of information fusion in pattern recognition is discussed from the angle of feature level fusion and decision level fusion.First, a new method of feature level fusion pattern recognition is presented. Feature Fusion Coefficients are defined to fuse the features extracted from different view of multiple sensors. By evaluating different feature fusion coefficients to different features, we can get the fusion feature of the pattern to be observed. Thus, feature level fusion pattern recognition can be realized.Second, in the theoretical framework of decision level fusion, we discuss the applications of D-S evidence theory in pattern recognition. In order to solve the question of obtaining the basic belief assignment of each evidence in the practical application of D-S theory in pattern recognition, we use neural network to obtain the basic belief assignments of the uncertainties in classification.Third, in the theoretical framework of decision level fusion, in order to solve the practical problems of fuzzy logic that the existing reasoning method is too simple in calculation to preserve lots of useful information, we apply D-S theory to fuzzy reasoning. In addition, we simulate the Attention Mechanism function of human brain to treat different rules in different ways.At last, in order to further explore the practicality of fusion pattern recognition, we use multimodal biometrics, which is based on fingerprint and hand geometry to validate the availability of the third method.更多还原