基于P-RBF神经网络的人脸识别算法研究

发布时间：2018-01-07 04:46

本文关键词：基于P-RBF神经网络的人脸识别算法研究　出处：《南昌航空大学》2017年硕士论文　论文类型：学位论文

【摘要】：数据预处理时,主成分分析算法(PCA算法)能够降低特征空间的维数。但是,由于涉及整体面部图像,使得在改变视点的情况下不能保证具有相同的识别率,故为了弥补局限性,在PCA算法的基础上,提出了线性判别分析算法(LDA算法)来提高处理不同类别图像时的识别率。本文首先阐述了由PCA与LDA相结合的新型算法,接着详细介绍了P-RBF神经网络的设计方法和具体实现过程,最后,在ATT数据库和耶鲁数据库中进行人脸识别试验,从而为P-RBF神经网络的人脸识别系统设计了一个最优的人脸识别方案。在本文中,提出了基于多项式的径向基函数神经网络(P-RBF NNS)作为主要识别部分的人脸识别系统。该系统有助于解决高维图像识别问题,其主要由图像数据预处理部分和图像数据识别部分构成。提出的P-RBF神经网络体系结构分为三个功能模块:条件部分、结论部分、和聚集部分。在P-RBF神经网络的条件部分,输入空间通过使用模糊C均值(FCM)算法来实现模糊聚类的分配。在P-RBF神经网络的结论部分中,使用如下三种多项式,如常数型、线性型和二次多项式型来作为连接函数。在P-RBF神经网络的聚集部分,通过采用模糊推理法获得P-RBF神经网络模型的系数。同时,将“如果-那么”规则作为该神经网络聚集部分的模糊规则集合。该神经网络的基本设计参数(包括学习速率,动量系数,模糊化系数和特征选择项)由差分进化(DE)算法进行优化。最后,在ATT数据库和耶鲁数据库进行人脸识别试验,实验结果表明,PCA-LDA结合算法具有更好的可行性和有效性,能有效实时的给出测试者的人脸识别结果。
[Abstract]:Data preprocessing, principal component analysis algorithm (PCA algorithm) can reduce the dimension of feature space. However, due to the whole face image, the change of viewpoint in case of guarantee has the same recognition rate, so in order to make up for the limitations, based on the PCA algorithm, presents a linear discriminant analysis algorithm (LDA to improve the recognition algorithm) when the rate of different categories of images. This paper describes the new algorithm by the combination of PCA and LDA, then introduces the design method of P-RBF neural network and realization process, finally, face recognition test in the ATT database and Yale database, so as to design a scheme for face recognition the optimal P-RBF neural network for face recognition system. In this paper, radial basis function neural network is proposed based on polynomial (P-RBF NNS) as the main part of the recognition of face recognition system System. The system is helpful to solve the high-dimensional image recognition problem, which is mainly composed of image data preprocessing and image data to identify parts. P-RBF neural network architecture proposed is divided into three functional modules: part, conclusion, and aggregation. In P-RBF neural network conditions, the input space by the use of fuzzy C means (FCM) algorithm to realize the distribution of fuzzy clustering. The P-RBF neural network in the conclusion part, use the following three kinds of polynomials, as usual number type, linear type and two polynomial as the connection function in the aggregation part of P-RBF neural network by using fuzzy inference method, coefficient of P-RBF neural network model. At the same time, the "If then" rules as the neural network fuzzy aggregation rules part of the collection. The basic design parameters of the neural network (including learning rate, momentum coefficient, fuzzy Feature selection coefficient) by differential evolution (DE) algorithm was optimized. Finally, face recognition test in the ATT database and Yale database, the experimental results show that the PCA-LDA algorithm has better feasibility and effectiveness of the face recognition results can effectively give the real-time testing.

【学位授予单位】：南昌航空大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】