摘要
随着网络层数的加深,卷积神经网络在训练过程中会出现梯度爆炸或者梯度消失的问题。为了解决这个问题,同时提高预测的精度,文中提出一种改进的残差网络。该网络主要对残差网络的结构进行改进,在直接映射层中将BN层和非线性激活层移动到卷积层之前,在恒等映射层中加入BN层和Conv2D层,这样不仅能解决加深网络情况下的梯度消失或梯度爆炸问题,而且预测的准确率和收敛速度也有很大的提升。通过实验证明,提出的改进残差网络算法相比于卷积神经网络准确率由之前的98.3%提升到1,损失值也从之前的0.26减少到0.024。该算法显著地提高了预测的准确度,降低了损失值,在流量分类中具有更好的应用。
With the deepening of the network layers, the problem of gradient explosion or gradient disappearance of convolutional neural networks will occur during the training process. In order to solve this problem and improve the accuracy of prediction, this paper proposes an improved residual network, which mainly improves the structure of the residual network. Moving the BN layer and the nonlinear activation layer to the front of the convolutional layer in the direct mapping layer, adding the BN layer and Conv2 D layer to the identity mapping layer.This not only solves the problem of gradient disappearance or gradient explosion in the case of deepening the network, but also greatly improves the accuracy and convergence speed. Comparing with the convolutional neural network’s accuracy, the experiments show that the accuracy of the proposed improved residual network algorithm has increased from 98.3% to 1, and the loss value has also been reduced from 0.26 to 0.024. This algorithm significantly improves the accuracy of prediction, reduces the loss value, and has better applications in traffic classification.
作者
严志兵
马自强
王恒
黄岩
YAN Zhibing;MA Ziqiang;WANG Heng;HUANG Yan(School of Information Engineering,Ningxia University,Yinchuan 750000,China)
出处
《黑龙江工程学院学报》
CAS
2022年第6期25-29,共5页
Journal of Heilongjiang Institute of Technology
基金
宁夏自然科学基金项目(2021AAC03114)。
关键词
卷积神经网络
流量分类
残差网络
恒等映射层
convolutional neural networks
traffic classification
improved residual networks
identity mapping layer
