Design and optimization of shared feature layer in deep convolutional neural networks
ZHANG Rendong
YE Shulin
HU Hongsheng
Abstract：As the depth of neural networks increases,the problems of the gradient vanishing and the gradient explosion become increasingly prominent.Meanwhile,improving feature reuse rate and reducing computational complexity have emerged as key challenges.This paper proposes a feature extraction network based on shared feature layer,and designs a dynamic iterative update mechanism.This mechanism integrates the output features of all previous layers and those of the current layer,and combines batch channel normalization to improve feature reuse efficiency.The shared feature layer is connected to the backbone network via residual connections,which enables effective feature updating and maintains gradient stability in the process of backpropagation.The experimental results show that the Top-1 error rate of SFLN-152 on ImageNet dataset is 19.00%,which is 11.34%and 5%lower than that of ResNet-152 and DPN-131,respectively.The classification error rates on the CIFAR-10 and CIFAR-100 datasets are 4.77%and 19.85%.
Keywords：feature extractiongradient stabilitybatch-channel normalizationshared feature layer
Publication Date：2025-09-30
Online Publishing Date：2025-10-15(First online date of this platform, not the publication date of the document)
Pages：7( 32-38 )