基于局部调节卷积神经网络的图像识别方法  

Image Recognition Method Based on Local Adjustment Convolutional Neural Network

作　　者：吴禄慎[1] 常参参 王晓辉[1] 陈华伟[1] WU Lushen;CHANG Cancan;WANG Xiaohui;CHEN Huawei(School of Mechanical and Electrical Engineering,Nanchang University,Nanchang 330031,China)

机构地区：[1]南昌大学机电工程学院,南昌330031

出　　处：《计算机工程》2018年第12期240-246,共7页Computer Engineering

基　　金：国家自然科学基金"汽车及飞机模具数字化快速修复与再制造技术"(51065021);国家自然科学基金"汽车模具表面缺陷的逆向建模及特征识别理论与算法"(51365037);南昌大学研究生创新专项资金"基于机器视觉的日用瓷器分拣系统"(cx2016065)

摘　　要：卷积神经网络在应用于图像识别时,会出现参数调节速度慢、迭代次数多以及小样本数据分类效果较差的问题。为此,提出一种局部调节卷积神经网络的方法。通过调节需求的大小,将参数对应的神经元分为关键部分和非关键部分,采用动态学习率和局部关键点修正,实现参数快速调节。在mnist、ORL、CIFAR-10和LFW上的实验结果表明,与DAP、UCNN等算法相比,该方法局部调节卷积神经网络的参数更快,在图像识别中达到识别精度需要的时间更少,而且整体识别率较高。Convolutional neural networks have some problems of slow parameter adjustment,more iterations,and poor classification of small sample data in the image recognition.Therefore,a method of locally adjusting a Convolutional Neural Network(CNN)is proposed.By adjusting the size of the demand,the neurons corresponding to the parameters are divided into key parts and non-critical parts,and dynamic learning rate and local oritical point correction are used to achieve rapid parameter adjustment.Experimental results on mnist,ORL,CIFAR-10 and LFW show that compared with DAP,UCNN and other algorithms,the local adjustment of the CNN parameters is faster.It has Less time and a higher overall recognition rate under the same recognition accuracy in image recognition.

关 键 词：深度学习 卷积神经网络 损失函数 局部调节 学习率 图像识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]