基于深度卷积神经网络的特征优化与分类识别方法研究

发布时间：2018-08-31 10:01

【摘要】：模式识别是人工智能领域内的一个重要分支,能够让机器观察周围环境并学习对不同的模式做出相应的判断。特征的好坏是模式识别系统发挥性能的关键,因此,需要一种通用的特征优化方法。本课题立足于模式识别问题中特征优化与分类识别两个方面,在使用深度卷积神经网络(Deep Convolutional Neural Network,DCNN)进行分析的基础上,研究在不同实际问题下如何对特征进行优化,从而对达到提高识别率的效果,并加深对于网络行为的理解。其主要研究内容包括:(1)在分析基于DCNN的二分类问题的特征优化方法的同时,结合脑机接口(Brain Computer Interface,BCI)中二分类运动想象的具体例子,提出基于DCNN的共同空间模式(Common Spatial Pattern,CSP)自适应特征优化方法。在原始脑电信号经过预处理的基础上,通过CSP空间变换获得其相应的特征矩阵。应用DCNN对特征矩阵进行学习,对收敛后的DCNN网络全连接层的权值进行分析,根据网络学习特性定义CSP矩阵特征筛选准则,得到降维高效的EEG特征集F,计算特征集F规模构建CNN分类器。我们工作在BCI2005Ⅳa竞赛数据集上进行实验测试,获得88.3%的识别准确率。本方法与CSP方法的改进方法s CSP和KLCSP方法在相同数据集上进行测试,平均识别准确率分别提升了3.2%和2.4%。(2)在二分类问题的基础上,进一步分析基于DCNN的多分类问题的特征优化方法的同时,针对语音信号认知中构建实际系统中需要优选有效特征的问题,提出了一种基于DCNN的特征降维方法。对原始语音情感数据提取大量特征,应用DCNN对特征矩阵进行学习,提取权值并根据网络学习特性定义特征筛选准则MCFR-DCNN,计算对比每类特征激活权值的不同,得到降维高效的语音情感认知特征集F。在中国科学院自动化研究所提供的多模态情感数据库CHEAVD上,我们提取全部八类情感数据进行了实验测试,使用全体特征集构建DCNN分类器相比基线类平均识别错误率减少了2.1%,而本方法得到的降维后特征集F通过相同的DCNN分类器相比基线类平均错误率较少了9.4%。本方法仅使用原特征集15%的特征,不仅减少了构筑实际语音情感识别系统的复杂程度,还使得识别错误率还有所减小。(3)在进一步分析使用DCNN对基于复杂特征的多分类问题进行研究的同时,结合具体的物体识别的例子,提出了基于DCNN的权值矩阵特性的复杂特征优化方法。使用DCNN对原始数据进行学习后提取全连接层权值,根据网络学习特性定义特征价值矩阵,进而将原始数据与特征价值矩阵进行线性组合得到新的数据集。在大规模图像数据库Image Net上进行了实验测试,分别使用3种分类器,相对于全部数据集,使用经过特征优化后的数据集在识别率上有所提高。本文研究针对不同数据的不同特征形式,采用DCNN网络学习特性进行特征二次优选与降维,为模式识别领域中的特征优化与分类识别问题提供了一个新的思路。
[Abstract]:Pattern recognition is an important branch in the field of artificial intelligence, which allows machines to observe the environment and learn to make corresponding judgments on different patterns. The quality of features is the key to the performance of pattern recognition system. Therefore, a general feature optimization method is needed. In two aspects of classification and recognition, based on the analysis of Deep Convolutional Neural Network (DCNN), this paper studies how to optimize the features under different practical problems, so as to improve the recognition rate and deepen the understanding of network behavior. The feature optimization method based on DCNN for binary classification problem is proposed. Combined with the specific examples of binary classification motion imagery in Brain Computer Interface (BCI), an adaptive feature optimization method based on DCNN Common Spatial Pattern (CSP) is proposed. The corresponding characteristic matrix is obtained by space transformation. DCNN is used to study the characteristic matrix, and the weights of the full connection layer of the convergent DCNN network are analyzed. According to the learning characteristics of the network, the feature selection criterion of CSP matrix is defined. The dimension-reduced and efficient EEG feature set F is obtained, and the scale of feature set F is calculated to construct the CNN classifier. We work in BCI 2005 IV. The improved CSP and KLCSP methods are tested on the same data set, and the average recognition accuracy is improved by 3.2% and 2.4%. (2) Based on the binary classification problem, the feature optimization of DCNN-based multi-classification problem is further analyzed. Meanwhile, a feature reduction method based on DCNN is proposed to optimize the effective features in speech recognition system. A large number of features are extracted from the original speech emotion data. DCNN is used to learn the feature matrix, extract the weights and define the feature selection criterion MCFR-DC according to the network learning characteristics. Based on the multi-modal emotion database CHEAVD provided by the Institute of Automation of the Chinese Academy of Sciences, we extracted all eight kinds of emotion data for experimental testing, and constructed DCNN classifier with all feature sets to compare the baseline class average. The recognition error rate is reduced by 2.1%, and the average error rate of feature set F is 9.4% less than that of baseline class by using the same DCNN classifier. This method uses only 15% features of the original feature set, which not only reduces the complexity of constructing actual speech emotion recognition system, but also reduces the recognition error rate. Further more, this paper analyzes the use of DCNN to study the multi-classification problem based on complex features, and proposes a complex feature optimization method based on DCNN weight matrix characteristics with specific object recognition examples. Matrix is used to combine the original data with the eigenvalue matrix linearly to get a new data set. Experiments are carried out on the large-scale image database Image Net. Three classifiers are used to improve the recognition rate of the optimized data set. With different feature forms, DCNN network learning characteristics are used to optimize and reduce the dimensions of features, which provides a new idea for feature optimization and classification in the field of pattern recognition.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP18

【参考文献】