基于改进Fisher准则的深度卷积神经网络识别算法  

Deep Convolution Neural Network Recognition Algorithm Based on Improved Fisher Criterion

作　　者：孙艳丰[1] 齐光磊[1] 胡永利[1] 赵璐[1] 

机构地区：[1]北京工业大学城市交通学院多媒体与智能软件技术北京市重点实验室,北京100124

出　　处：《北京工业大学学报》2015年第6期835-841,共7页Journal of Beijing University of Technology

基　　金：国家自然科学基金资助项目(61370119)

摘　　要：为了有效利用深度学习技术自动提取特征的能力,并解决当训练样本量减少或者迭代次数降低时识别性能急速下降的问题,提出了基于Fisher准则的深度学习算法.该方法在前馈传播时,采用卷积神经网络自动提取图像的结构信息等特征,同时利用卷积网络共享权值和池化、下采样等方法减少了权值个数,降低了模型复杂度;在反向传播权值调整时,采用了基于Fisher的约束准则.在权值的迭代调整时既考虑误差的最小化,又同时让样本保持类内距离小,类间距离大,从而使权值能更加快速地逼近有利于分类的最优值,当样本量不足或训练迭代次数不多时可有效地提高系统的识别率.大量的实验结果证明:该基于Fisher准则的混合深度学习算法在标签样本不足或者较少训练次数的情况下依然能达到较好的识别效果.To effectively make use of deep learning technology automatic feature extraction ability, and solve the problem when the training sample size reduced or the iteration times reduced the recognition performance fell sharply, this paper proposed a deep learning algorithm based on Fisher criterion. In the feed forward spread, this method used convolution neural network to extract automatically image features such as structural information, and used convolution network of sharing weights and pooling, sub-sampling methods to reduce the weight number, and the method reduced the model complexity. When the back propagation adjusted the weights, it adopted the constraints based on Fisher criterion. At the same time, it kept the samples in small distance with-class and large distance between-class, so that the weights could be more close the optimal value for classification. It improved the recognition rate effectively when the sample size was insufficient or when it had few training iterations. A large number of experiments show that when the label samples are insufficient and the training iteration fewer, the hybrid deep learning algorithm based on Fisher criterion still achieves good recognition effect.

关 键 词：深度学习 卷积神经网络 FISHER准则 反向传播(BP)算法 人脸识别 手写字识别 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]