【作者】 吴浩；

【导师】 金立左；

【作者基本信息】 东南大学 ， 模式识别与智能系统， 2018， 硕士

【摘要】 模式分类是人工智能的基本问题,在国防、安保、制造、交通、金融、环境、医疗等众多领域都有着广泛应用价值。尤其近年来,随着社会生产与民生服务的智能化大发展,受到越来越多关注。模式分类是指通过构造分类模型将数据映射到某个给定类别。解决模式分类问题一般从特征和分类两个角度着手。特征是训练分类器的基础,与分类算法共同影响分类结果的好坏。Adaboost是一种具有特征选择能力的分类算法,具有分类准确、计算快速的优点,为此,论文针对Adaboost算法,重点研究模式分类的特征选择与分类算法设计问题。特征提取问题。研究了Haar特征和Gabor特征的提取原理和算法实现,从计算复杂性等方面,对比分析了这两类特征的优缺点,从实际应用需求出发,将Haar特征用于二类识别问题,将Gabor特征用于多类识别问题。二类识别问题。研究了Adaboost算法原理,分析了Adaboost算法的特征选择机理。介绍了目标检测问题中正负样本不均衡问题,研究了如何利用硬级联分类器检测目标。为了克服硬级联结构的局限性,论文重点研究了柔性级联结构,提出了新的分类器训练算法,涉及特征选择与阈值整定。特征选择通过候选集扫描优化采样过程,阈值整定通过直接反向修剪算法优化速度。采用头肩图像集和车牌图像集进行目标检测实验,对比分析硬级联分类器与柔性级联分类器的性能,结果表明柔性级联分类器在检测精度与计算速度上具有优势。多类识别问题。针对一对多的分类方式,研究了将Adaboost二类分类器组合为多类分类器的GAB.OVA算法。分析了GAB.OVA算法中的预测模型,针对其特征数随着类别数线性增长的局限性,引入了复合树桩弱分类器,提出了其用于多分类问题的BCS算法。该算法实现了类间特征共享,消除了特征数与类别数的线性增长关系,降低了模型复杂度。采用MNIST手写字符数据集与车牌字符数据集进行目标分类实验,结果表明BCS算法在分类精度和计算速度上具有优势。更多还原

【Abstract】 Pattern classification is the basic problem of artificial intelligence,which has been widely used in many fields such as defense,security,manufacturing,transportation,finance,environment,and medical care.In recent years,it has been paid more and more attention due to development of social production and livelihood services.Pattern classification is to mapping data to a given category by constructing a classification model.Pattern classification problems are generally solved from aspects of features and classification.Feature is the basis for training the classifier,which affects models’ performance together with classification algorithm.Adaboost is a classification algorithm with feature selection ability,whose advantages are high accuracy and fast speed.Therefore,the paper focuses on the feature selection and classification algorithm design in pattern classification problems based on Adaboost.Feature extraction.The main ideas and algorithm implementation of Haar features and Gabor features are studied.From the aspects of computational cost and so on,the advantages and disadvantages of these two kinds of features are compared and analyzed.Considering practical application requirements,Haar features are used for the two-class recognition problem while Gabor features is used for multi-class recognition problems.Two-class recognition problem.The basic idea of Adaboost algorithm is studied and the feature selection mechanism is then analyzed.To deal with the imbalance between positive and negative samples in object detection problem,it is studied that how to use the hard-cascade classifier to detect objects.To overcome the limitations of the hard-cascade structure,the paper focuses on the soft-cascade structure and proposes a new classifier training algorithm,which involves feature selection and threshold calibration.In feature selection,candidate set scanning is applied to optimizes the sampling process.In threshold calibration,direct back pruning is applied to improve predict speed.Head-shoulder image sets and license plate image sets are used for target detection experiments.The performance of the hard-cascade classifier and the soft-cascade classifier was compared and analyzed.The results show that the soft-cascade classifier performs better in both accuracy and speed.Multi-class recognition problem.Based on the one-vs-all classification mode,the GAB.OVA algorithm is studied to explain how to combine the two-class classifiers into a multi-class classifier.The prediction model of GAB.OVA algorithm is analyzed.To overcome the limitation that the number of features increases linearly with the number of categories,a weak classifier called composite stumps is introduced,and a new multi-classification algorithm BCS is proposed.The BCS algorithm applies feature sharing among classes,eliminates the linear relationship between numbers of features and categories,and so reduces the complexity of prediction model.MNIST handwritten character datasets and license plate character datasets is used for multi-class recognition experiments,the results show that BCS algorithm has lower computation cost and no decrease in accuracy.更多还原