【作者】 赵仲秋；

【导师】 黄德双；

【作者基本信息】 中国科学技术大学 ， 模式识别与智能系统， 2007， 博士

【摘要】 目前，模式识别已经发展了很多分类方法。然而，随着应用领域的不断推广和处理数据的不断庞大，模式识别系统的学习速度、分类精度和代价不断受到挑战。特别是在实际应用过程中往往会出现一些特殊的复杂问题，比如大样本集问题、训练集类别不对称问题、有类别标记样本少而无类别标记样本较多的问题等等。对于这些特殊问题，已有的模式识别方法已经不能够满足实际应用中对识别精度和学习速度的要求，因此，有必要深入探讨和研究针对这些特殊问题的特定的分类模型。本论文主要针对模块神经网络解决大样本集学习问题、训练集中类别不对称问题、以及流形学习应用于半监督分类等课题来展开全面而系统的研究。研究成果丰富了针对这些特殊的复杂问题的分类模型，提高了它们的分类效率。全文的主要工作体现在以下几个方面：1、提出了一种基于新的任务分解技术的矩阵模块神绎网络分类系统，它将一个复杂分类任务分解为多个简单的子任务来解决，每个子任务只是在两个子空间内进行，且由一个具有简单结构的神经网络模块来完成，所有网络模块将组成一个神绎网络矩阵，最终将该神经网络矩阵的输出矩阵集成得到最终分类结果。通过理论分析和模拟实验证明，该矩阵模块神经网络能节省学刊时间，提高分类精度。2、成功将矩阵模块神经网络应用于人脸和掌纹识别系统。对于掌纹识别问题，提出了一种有效的2DPCA(w／o3)+PCA特征抽取技术，此特征提取方法比其它特征提取方法花费较少的抽取时间，却能取得更好的分类精度。3、提出用矩阵模块神经网络来解决非对称模式分类问题的模型结构。它将非对称模式分类问题分解为一系列对称的两类问题来解决，每个两类问题由一个结构简单的网络来解决，并且仅使用简单的网络学习算法就能够取得较好的分类结果。该矩阵模块神经网络能有效地减少非对称分类问题的学习时间，提高其分类精度。4、提出一种改进的基于黎曼流形和最小误差类别映射的半监督学习算法，使其能直接应用于多类半监督学习问题。该改进算法在保持与原算法相同的分类精度的基础上，能够大大提高学习速度。5、提出了一种光谱映射的改进算法——半监督光谱映射用于半监督分类，取得了较好的分类效果。该算法在映射时添加了类别信息，并且用沿着流形表而的测地距离取代了原来的欧氏距离作为样本点之间差异性的测度。该改进算法提高了映射的性能，并且取得了较好的分类结果。更多还原

【Abstract】 At present, there have been many classification techniques well developed in pattern recognition field. However, the broadening of fields and the enlarging of sizes of datasets dealt with in real applications are challenging the learning speeds and classification accuracies of all sorts of pattern classification systems. Specially, some special complex cases such as the classification problems with large size of training set, unbalanced training set and partially labeled training set demand designing some special and more effective classification models. Therefore, it is necessary for us to thoroughly investigate the classification models so as to solve those highly complex problems. This thesis is focused on comprehensively and systemically solving the classification problems with large size of training set and imbalanced training set by using modular neural networks, as well as semi-supervised classification problems by using manifold learning. The obtained results enrich and perfect the classification models and enhance the classification performance for these complex problems. The main works in the thesis can be stated as follows:1. A classification structure of matrix modular neural network based on a novel task decomposition technique was proposed, which can decompose a complex task into several easier subtasks between subspace pairs. Each subtask is then solved by a simple perceptron. All of these perceptron modules form a perceptron matrix structure, which produces a matrix of outputs that will be fed to an integration machine so that a classification decision can be efficiently made. This method can greatly speed up training of neural networks and obviously enhance the generalization capability for distinguishing unknown samples according to our experiments and theoretical analyses.2. The proposed matrix modular neural network was successfully applied into face recognition and palmprint recognition. Furthermore, for palmprint recognition, a feature extraction technique called ’2DPCA(w/o3)+PCA’, which consumes less eatraction time but obtains better classification performance than other feature extraction techniques, was also proposed.3. A structure of matrix modular neural network was proposed to deal with the imbalanced pattern classification problems. By this matrix modular neural network, an imbalanced classification problem can be transformed into a set of symmetrical two-class problems, each of which can be solved easily by a simple network. The experimental results showed that the matrix modular neural network could reduce the CPU consumption for the training, and also improve the classification performance.4. A modified version to semi-supervised learning algorithm based on Riemannian manifold and mapping for minimum error sum was proposed to make it applicable to multi-classes semi-supervised learning. The modified algorithm largely increases the learning speed, and at the same time attains the satisfying classification performance, which is not lower than that of the original algorithm.5. An improved version to spectral mapping, referred to as semi-supervised spectral mapping, was proposed to implement semi-supervised learning. This new method adds the label information into the mapping process, and adopts the geodesic distance rather than Euclidean distance as the measure of the difference between two data points. The experimental results show that the proposed method yields significant benefits for partially labeled classification with respect to the previous methods.更多还原