基于改进Fisher判别准则的卷积神经网络设计  

Design of Convolutional Neural Network Based onImproved Fisher Discriminant Criterion

作　　者：徐小雨 赵龙章 程晓月 何志超 XU Xiaoyu;ZHAO Longzhang;CHENG Xiaoyue;HE Zhichao(College of Electrical Engineering and Control Science,Nanjing Tech University,Nanjing 211816,China)

机构地区：[1]南京工业大学电气工程与控制科学学院,南京211816

出　　处：《计算机工程》2020年第11期255-260,266,共7页Computer Engineering

基　　金：国家自然科学基金(61403189)。

摘　　要：传统卷积神经网络(CNN)在建模过程中由于数据样本量不足容易出现过拟合现象,且对随机数据泛化能力较差。为此,设计一种结合改进Fisher判别准则与GRV模块的卷积神经网络(FDCNN)。使用CNN学习从输入图像到多维欧式空间的映射关系,采用基于改进Fisher判别准则的损失函数进行网络模型训练并将人脸样本数据投影到低维空间,保证类内离散度尽量小的同时类间离散度尽量大以达到最佳人脸分类效果。引入融合GoogleNet、ResNet和VGGNet网络结构特点的GRV模块,提高CNN网络表达能力并降低网络模型复杂度。实验结果表明,当训练样本数量为840时,FDCNN模型在CBCL数据集上的识别率为93.4%,相比传统CNN模型、基于改进Fisher判别准则的全连接神经网络模型等网络模型识别率更高且泛化能力更好。Traditional Convolutional Neural Network(CNN)are prone to overfitting,and have performance in generalizing random data.To address the problem,this paper designs a CNN based on Fisher discriminant criterion and GRV module,called FDCNN.The method uses a loss function based on improved Fisher discriminant criterion to train the model,and maps the sample data of human faces to low-dimensional space.So the dispersion of the same type of faces in the mapping space is minimized while the dispersion of different types of faces is maximized to achieve the optimal face classification performance.In addition,this paper combines the advantages of GoogleNet,ResNet,VGGNet network structures to design a new GRV module,which improves the representation ability of the CNN and simplifies the network model.Experimental results show that when the number of training samples is 840,the recognition rate of the proposed FDCNN model on the CBCL dataset reaches 93.4%,which outperforms the traditional CNN model,fully connected neural network model based on improved Fisher discriminant criterion,and other network models.Also,the FDCNN model has better generalization ability than the above models.

关 键 词：卷积神经网络 FISHER判别准则 损失函数 离散度 泛化能力 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]