基于特征图切分的轻量级卷积神经网络  

A Lightweight Convolutional Neural Network Architecture with Slice Feature Map

作　　者：张雨丰 郑忠龙[1] 刘华文[1] 向道红 何小卫[1] 李知菲[1] 何依然 KHODJA Abd Erraouf ZHANG Yufeng;ZHENG Zhonglong;LIU Huawen;XIANG Daohong;HE Xiaowei;LI Zhifei;HE Yiran;KHODJA Abd Erraouf(Department of Computer Science, Zhejiang Normal University, Jinhua 321004;Department of Mathematics, Zhejiang Normal University, Jinhua 321004)

机构地区：[1]浙江师范大学计算机科学与工程系,金华321004 [2]浙江师范大学数学系,金华321004

出　　处：《模式识别与人工智能》2019年第3期237-246,共10页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金项目(No.61672467;61572443;11871438)资助~~

摘　　要：卷积神经网络模型所需的存储容量和计算资源远超出移动和嵌入式设备的承载量,因此文中提出轻量级卷积神经网络架构(SFNet).SFNet架构引入切分模块的概念,通过将网络的输出特征图进行"切分"处理,每个特征图片段分别输送给不同大小的卷积核进行卷积运算,将运算得到的特征图拼接后由大小为1×1的卷积核进行通道融合.实验表明,相比目前通用的轻量级卷积神经网络,在卷积核数目及输入特征图通道数相同时,SFNet的参数和计算量更少,分类正确率更高.相比标准卷积,在网络复杂度大幅降低的情况下,切分模块的分类正确率持平甚至更高.The capacities of mobile and embedded devices are quite inadequate for the requirement of the storage capacity and computational resources of convolutional neural network models. Therefore, a lightweight convolutional neural network architecture, network with slice feature map, named SFNet, is proposed. The concept of slice block is introduced. By performing the “slice” processing on the output feature map of the network, each feature map segment is respectively sent to a convolution kernel of different sizes for convolution operation, and then the obtained feature map is concatenated. A simple 1×1 convolution is utilized to fuse the channels of the feature map. The experiments show that compared with the state-of-the-art lightweight convolutional neural networks, SFNet has fewer parameters and floating-point operations, and higher classification accuracy with the same number of convolution kernels and input feature map channels. Compared with the standard convolution, in the case of a significant reduction in network complexity, the classification accuracy is same or higher.

关 键 词：卷积神经网络 轻量级网络 切分模块 特征图切分 组卷积 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]