一种嵌入并行通道蓝图分离卷积的图像分类算法  

Blueprint Separation Convolution is Embedded in Parallel Channels for Image Classification

作　　者：邱飞岳[1,2] 孔德伟 张志勇 章国道 QIU Fei-yue;KONG De-wei;ZHANG Zhi-yong;ZHANG Guo-dao(College of Education Science and Technology,Zhejiang University of Technology,Hangzhou 310023,China;College of Computer Science and Technology,Zhejiang University of Technology,Hangzhou 310023,China;Information Engineering,East China University of Technology,Nanchang 330013,China)

机构地区：[1]浙江工业大学教育科学与技术学院,杭州310023 [2]浙江工业大学计算机科学与技术学院,杭州310023 [3]东华理工大学信息工程学院,南昌330013

出　　处：《小型微型计算机系统》2021年第12期2592-2599,共8页Journal of Chinese Computer Systems

基　　金：浙江省重点研发计划基金资助项目(2018C01080)资助。

摘　　要：卷积神经网络计算复杂度高,占用资源多,使得其很难部署到移动嵌入式设备上.为了解决这类限制,本文提出了一种基于嵌入并行通道蓝图分离卷积(B2SENet)模型的图像分类算法,首先通过引入并行SENet通道,使得不同内核大小的并行分支融合在一起,增强全局感受野,其次在并行SE通道中改变传统卷积操作方式,适配蓝图分离卷积(BS),有效减少卷积层参数以及模型体积.最后在并行通道融合过程中引入soft attention机制,获取通道权重系数,使得模型强化对特征的选择性,在CIFAR10和100以及ImageNet2012基准数据集上对比,实验表明B2SENet(BS+并行通道SENet)在模型复杂度以及准确率上均优于MobileNet和ShuffleNet,使得其在嵌入式设备网络模型中表现相当出色.The convolutional neural network has high computational complexity that takes up a lot of resources,making it difficult to deploy to mobile embedded devices. In order to solve this kind of limit,this paper puts forward a kind of based on Blueprint Separation Convolution is Embedded in Parallel Channels( B2 SENet) for Image Classification,Firstly by introducing parallel SENet channel,makes different parallel branches fuses in together,the size of the kernel enhanced global receptive field,Secondly,the traditional convolution operation mode is changed with parallel SE channels,adapter blueprint separation convolution( BS),effectively reduce convolution layer parameters of the model and volume. Finally introduced attention in the process of parallel channel integration mechanism,gain weight coefficients of the channel,strengthen the modeling on characteristics of selective,in 100 and ImageNet2012 CIFAR10 and benchmark data sets on the contrast,experiments show that BS2 ENet( BS + Parallel channels SENet) on the model complexity and accuracy is superior to the existing most advanced models,its excellent performance in the embedded devices network model.

关 键 词：卷积神经网络 图像分类 残差网络 蓝图分离卷积 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]