摘要
针对传统ResNet网络存在丢失图像有用信息以及参数冗余等问题,论文提出一种改进ResNet的A-ResNet模型。引入有shortcut连接的残差注意力模块,增强对目标对象的关注度;引入Dropout层,防止过拟合现象,提升识别精度;调整网络架构,加快训练收敛速度及提高识别精度。实验结果表明,A-ResNet模型相比传统ResNet网络实现约2%的top-1精度的提高。
Aiming at the problems of lost useful information and parameter redundancy in traditional ResNet networks,an im⁃proved A-ResNet model of ResNet is proposed.A residual attention module with a shortcut connection is introduced to increase the focus on the target object,and the Dropout layer is introduced to prevent overfitting,which improves recognition accuracy,and the network architecture is adjusted to speed up the training convergence and the recognition accuracy is improved.The experimen⁃tal results show that the A-ResNet model achieves an improvement of about 2%of top-1 accuracy compared to the traditional ResNet network.
作者
郑秋梅
谭丹
王风华
ZHENG Qiumei;TAN Dan;WANG Fenghua(Department of Computer and Communication Engineering,China University of Petroleum,Qingdao 266580)
出处
《计算机与数字工程》
2021年第5期947-951,965,共6页
Computer & Digital Engineering
