摘要
结构参数量和计算量限制了卷积神经网络在移动设备上的应用。主要研究在尽量保持精度的前提下减少结构参数量和计算量。针对分组卷积引起的分组通道间不流通的问题,提出分组瓶颈;针对如何提升分类精度问题,提出奇异瓶颈;使用上述策略改进SqueezeNet,提出轻量化结构SlimNet。实验表明:引入分组瓶颈和奇异瓶颈具有有效性,提出的轻量化结构Slim Net在分类精度、结构参数量及计算量上均优于SqueezeNet。
Structure parameters and computations limit the application of convolutional neural networks(CNNs)in mobile devices.We mainly studied how to reduce the amount of structural parameters and computations while keeping the accuracy as far as possible.The grouped bottleneck was proposed according to the interchannel congestion caused by grouped convolution,singular bottleneck was proposed to improve classification accuracy,and the light-weight structure,SlimNet,was proposed after improving SqueezeNet using the above strategies.The experiment results demonstrate the effectiveness of grouped bottleneck and singular bottleneck SlimNet is superior to SqueezeNet in terms of classification accuracy,structural parameters and computation.
作者
董艺威
于津
Dong Yiwei;Yu Jin(Department of Computer Science and Technology,College of Engineering,Shantou University,Shantou 515063,Guangdong,China)
出处
《计算机应用与软件》
北大核心
2018年第11期226-232,共7页
Computer Applications and Software
关键词
图像分类
卷积神经网络
轻量化
分组卷积
分组瓶颈
奇异瓶颈
SlimNet
Image classification
Convolutional neural network
Light-weight
Group convolution
Grouped bottleneck
Singular bottleneck
SlimNet
