面向移动端图像分类的轻量级CNN优化  

Lightweight CNN optimization for mobile image classification

作　　者：张晓青 刘小舟 陈登[3,4] ZHANG Xiao-qing;LIU Xiao-zhou;CHEN Deng(College of Computer Science and Technology,Zhejiang University of Technology,Hangzhou 310023,China;Health Information Center of Zhejiang Province,Hangzhou 310006,China;Zhejiang Academy of Science and Technology Information,Hangzhou 310006,China;Key Laboratory of Open Data Zhejiang Province,Hangzhou 310053,China)

机构地区：[1]浙江工业大学计算机科学与技术学院,浙江杭州310023 [2]浙江省卫生健康信息中心,浙江杭州310006 [3]浙江省科技信息研究院,浙江杭州310006 [4]浙江省数据开放重点实验室,浙江杭州310053

出　　处：《计算机工程与设计》2024年第2期436-442,共7页Computer Engineering and Design

基　　金：浙江省科技厅省重点研发基金项目(2022C01083);国家自然科学基金项目(J2124006)。

摘　　要：为解决图像分类算法由于计算量大和参数冗余难以应用在存储空间与计算能力受限的移动设备上的问题,提出一种轻量的卷积计算模块Extremely Lightweight Block(ELBlock),采用逐点卷积叠加深度可分离卷积的方法,对逐点卷积进行分组,增加相邻层过滤器之间的对角相关性,进一步降低卷积操作的计算复杂度;利用通道混洗关联输入和输出通道,提高特征的信息表达能力;基于ELBlock设计一个极其轻量的小型神经网络架构ELNet,结构更加简洁、高效。在Android手机上的实验结果表明,所提ELNet在保证分类精度的同时,具有计算量小、参数少和推理时间短的优点。To solve the problem that the image classification algorithm is difficult to be applied to mobile devices with limited sto-rage space and computing power due to the large amount of calculation and the parameter redundancy,a lightweight convolution calculation module,namely Extremely Lightweight Block(ELBlock)was proposed.The method of point-by-point convolution superposition depthwise separable convolution was adopted.The point-by-point convolution was grouped to increase the diagonal correlation between filters of adjacent layers and the computational complexity of convolution operation was further reduced.The channel shuffle was used to correlate the input and output channels to improve the information expression ability of features.Based on ELBlock,an extremely lightweight small neural network architecture ELNet was designed,which was more concise and efficient.Experimental results on Android mobile phones show that the proposed ELNet has the advantages of less computation,fewer parameters and shorter inference time while ensuring the prediction accuracy.

关 键 词：深度学习 图像分类 轻量级神经网络 模型优化 模型压缩 模型部署 移动终端 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]