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。

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

Abstract:

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.

Key words: Deep Learning(DL), Convolutional Neural Network(CNN), loss function, local adjustment, learning rate, image recognition

中图分类号: 

• TP391