紧凑的神经网络模型设计研究综述

doi:10.3778/j.issn.1673-9418.1912079

摘要/Abstract

摘要：

近年来卷积神经网络在广泛的应用中取得了优秀的表现，但巨大的资源消耗量使得其应用于移动端和嵌入式设备成为了挑战。为了解决此类问题，需要对网络模型在大小、速度和准确度方面做出平衡。首先，从模型是否预先训练角度，简要介绍了网络压缩与加速的两类方法——神经网络压缩和紧凑的神经网络。具体地，阐述了紧凑的神经网络设计方法，展示了其中不同运算方式，强调了这些运算的特点，并根据基础运算的不同，将其分为基于空间卷积的模型设计和基于移位卷积的模型设计两大类，然后每类分别选取三个网络模型从基础运算单元、核心构建块和整体网络结构进行论述。同时，分析了各网络以及常规网络在ImageNet数据集上的性能。最后，总结了现有的紧凑神经网络设计技巧，并展望了未来的发展方向。

关键词: 卷积神经网络（CNN）, 轻量化, 移位操作, 卷积方式

Abstract:

In recent years, convolutional neural network has achieved excellent performance in a wide range of applications, but it consumes huge resources, which is a challenge for its application to mobile terminals and embedded devices. To this end, network models need to be balanced in size, speed and accuracy. Firstly, two methods of network compression and acceleration, neural network compression and compact neural network, are briefly intro-duced from the perspective of whether the model is pre-trained or not. Specifically, this paper describes the design method of compact neural network, shows the different operation modes, emphasizes the characteristics of these operations, and divides them into two categories: model design based on spatial convolution and model design based on shift convolution according to the different basic operations. Then, each class selects three network models to discuss from the basic operation unit, the core building block and the overall network structure. At the same time, the performance of each network and the conventional network on the ImageNet dataset is analyzed. Finally, this paper summarizes the existing design techniques of compact neural network and looks forward to the future develop-ment direction.

Key words: convolutional neural network (CNN), lightweight, shift operation, convolution method

郎磊，夏应清. 紧凑的神经网络模型设计研究综述[J]. 计算机科学与探索, 2020, 14(9): 1456-1470.

LANG Lei, XIA Yingqing. Survey on Compact Neural Network Model Design[J]. Journal of Frontiers of Computer Science and Technology, 2020, 14(9): 1456-1470.

参考文献

[1] Liu G X, Liu S Y, Wu J F, et al. Machine vision object detec-tion algorithm based on deep learning and application in bank-note detection[J]. China Measurement & Test, 2019, 45(5): 1-9.刘桂雄, 刘思洋, 吴俊芳, 等. 基于深度学习的机器视觉目标检测算法及在票据检测中应用[J]. 中国测试, 2019, 45(5): 1-9.
[2] Nanjiagelie, Li R, Wang H X, et al. Ultrasound image classi-fication of hepatic echinococcosis using deep learning[J].Journal of Shenzhen University (Science and Engineering),2019, 36(6): 702-708.南嘉格列, 李锐, 王海霞, 等. 基于深度学习的肝包虫病超声图像分型研究[J]. 深圳大学学报(理工版), 2019, 36(6):702-708.
[3] Harikrishnan J, Sudarsan A, Sadashiv A, et al. Vision-face recognition attendance monitoring system for surveillance using deep learning technology and computer vision[C]//Pro-ceedings of the 2019 International Conference on Vision To-wards Emerging Trends in Communication and Networking, Vellore, March 30-31, 2019. Piscataway: IEEE, 2019: 1-5.
[4] Sampedro C, Rodriguez-Vazquez J, Rodriguez-Ramos A, et al. Deep learning-based system for automatic recognition and diagnosis of electrical insulator strings[J]. IEEE Access, 2017, 7: 101283-101308.
[5] Zhang L, Yuan F N, Zhang W R, et al. Review of fully con-volutional neural network[J]. Computer Engineering and App-lications, 2020, 56(1): 25-37.章琳, 袁非牛, 张文睿, 等.?全卷积神经网络研究综述[J].计算机工程与应用,2020,56(1):25-37.
[6] Li H, Wan X X. Image style transfer algorithm under deep con-volutional neural network[J]. Computer Engineering and App-lications, 2020, 56(2): 176-183.李慧, 万晓霞.?深度卷积神经网络下的图像风格迁移算法[J]. 计算机工程与应用, 2020,56(2):176-183.
[7] Mao B C, Chen S K, Xie Y, et al. Exploration of classical deep learning algorithm in intelligent classification of Chinese randomized controlled trials[J]. Chinese Journal of Evidence-Based Medicine, 2019, 19(11): 1262-1267.毛渤淳, 陈圣恺, 谢雨, 等. 经典深度学习算法对中文随机对照试验智能判别应用[J]. 中国循证医学杂志, 2019, 19(11): 1262-1267.
[8] Wang X, Li C, Chen J. Dilated convolution neural networks for Chinese word segmentation[J]. Journal of Chinese Infor-mation Processing, 2019, 33(9): 24-30.王星, 李超, 陈吉. 基于膨胀卷积神经网络模型的中文分词方法[J]. 中文信息学报, 2019, 33(9): 24-30.
[9] Jaf S, Calder C. Deep learning for natural language parsing[J]. IEEE Access, 2019, 7: 131363-131373.
[10] Mahmud M, Shamim Kaiser M, Hussain A, et al. Appli-cations of deep learning and reinforcement learning to biolo-gical data[J]. IEEE Transactions on Neural Networks and Learning Systems, 2018, 29(6): 2063-2079.
[11] Feng M C, Zheng J B, Ren J C, et al. Big data analytics and mining for effective visualization and trends forecasting of crime data[J]. IEEE Access, 2019, 7: 106111-106123.
[12] Ye Z Y, Feng A M, Gao H. Customer purchasing power prediction of Google store based on deep LightGBM ense-mble learning model[J]. Journal of Computer Applications,2019, 39(12): 3434-3439.叶志宇, 冯爱民, 高航. 基于深度LightGBM集成学习模型的谷歌商店顾客购买力预测[J]. 计算机应用, 2019, 39(12): 3434-3439.
[13] Morgado P, Vasconcelos N. Semantically consistent regula-rization for zero-shot recognition[C]//Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recog-nition, Honolulu, Jul 21-26, 2017. Washington: IEEE Com-puter Society, 2017: 2037-2046.
[14] Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition[J]. arXiv:1409.1556, 2014.
[15] Szegedy C, Liu W, Jia Y Q, et al. Going deeper with convolu-tions[C]//Proceedings of the 2015 IEEE Conference on Com-puter Vision and Pattern Recognition, Boston, Jun 7-12, 2015. Piscataway: IEEE, 2015: 1-9.
[16] Wang K J, Zhao Y D, Xing X L. Deep learning in driverless vehicles[J]. CAAI Transactions on Intelligent Systems, 2018, 13(1): 55-69.王科俊, 赵彦东, 邢向磊. 深度学习在无人驾驶汽车领域应用的研究进展[J]. 智能系统学报, 2018, 13(1): 55-69.
[17] Lai J, Rao R. Application of deep reinforcement learning in indoor UAV target search[J]. Computer Engineering and App-lications, 2020, 56(17): 156-160.赖俊, 饶瑞. 深度强化学习在室内无人机目标搜索中的应用[J]. 计算机工程与应用, 2020, 56(17): 156-160.
[18] Kwon M, Ju M, Choi S. Classification of various daily be-haviors using deep learning and smart watch[C]//Proceed-ings of the 2017 International Conference on Ubiquitous and Future Networks, Milan, Jul 4-7, 2017. Piscataway: IEEE, 2017: 735-740.
[19] Khan S, Javed M H, Ahmed E, et al. Facial recognition using convolutional neural networks and implementation on smart glasses[C]//Proceedings of the 2019 International Conference on Information Science and Communication Technology, Karachi, Mar 9-10, 2019. Piscataway: IEEE, 2019: 1-6.
[20] Chen C J, Chen K C, Martin-Kuo M. Acceleration of neural network model execution on embedded systems[C]//Proceed-ings of the 2018 International Symposium on VLSI Design, Automation and Test, Hsinchu, China, Apr 16-19, 2018. Piscat-away: IEEE, 2018: 1-3.
[21] Wu J, Qian X Z. Compact deep convolutional neural network in image recognition[J]. Journal of Frontiers of Computer Science and Technology, 2019, 13(2): 275-284.吴进, 钱雪忠.紧凑型深度卷积神经网络在图像识别中的应用[J].计算机科学与探索,2019,13(2):275-284.
[22] Li Q H, Li C P, Zhang J, et al. Survey of compressed deep neural network[J]. Computer Science, 2019, 46(9): 1-14.李青华, 李翠平, 张静, 等. 深度神经网络压缩综述[J]. 计算机科学, 2019, 46(9): 1-14.
[23] Ji R R, Lin S H, Chao F, et al. Deep neural network com-pression and acceleration: a review[J]. Journal of Computer Research and Development, 2018, 55(9): 1871-1888.纪荣嵘, 林绍辉, 晁飞, 等. 深度神经网络压缩与加速综述[J]. 计算机研究与发展, 2018, 55(9): 1871-1888.
[24] Han S, Mao H, Dally W J. Deep compression: compressing deep neural networks with pruning, trained quantization and Huffman coding[J]. arXiv:1510.00149, 2015.
[25] Huizi M, Han S, Pool J, et al. Exploring the regularity of sparse structure in convolutional neural networks[J]. arXiv:1705.08922, 2017.
[26] Wen W, Wu C, Wang Y, et al. Learning structured sparsity in deep neural networks[C]//Proceedings of the Advances in Neural Information Processing Systems. Berlin: Springer,2016: 2074-2082.
[27] Liu Z, Li J G, Shen Z Q, et al. Learning efficient convolu-tional networks through network slimming[C]//Proceedings of the 2017 IEEE International Conference on Computer Vision, Venice, Oct 22-29, 2017. Washington: IEEE Computer Society, 2017: 2755-2763.
[28] Luo J H, Wu J X, Lin W Y. ThiNet: a filter level pruning method for deep neural network compression[C]//Proceed-ings of the 2017 IEEE International Conference on Computer Vision, Venice, Oct 22-29, 2017. Washington: IEEE Computer Society, 2017: 5068-5076.
[29] Zhang X, Zou J, He K, et al. Accelerating very deep convolu-tional networks for classification and detection[J]. IEEE Trans-actions on Pattern Analysis and Machine Intelligence, 2016, 38(10): 1943-1955.
[30] Lebedev V, Ganin Y, Rakhuba M, et al. Speeding-up convolu-tional neural networks using fine-tuned CP-decomposition[EB/OL]. [2019-10-09]. https://arxiv.org/pdf/1412.6553.pdf.
[31] Wu J X, Leng C, Wang Y H, et al. Quantized convolutional neural networks for mobile devices[C]//Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recog-nition, Las Vegas, Jun 27-30, 2016. Washington: IEEE Com-puter Society, 2016: 4820-4828.
[32] Hu Q H, Wang P S, Cheng J. From Hashing to CNNs: training binary weight networks via Hashing[J]. arXiv:1802.02733,2018.
[33] Zagoruyko S, Komodakis N. Paying more attention to atten-tion: improving the performance of convolutional neural networks via attention transfer[J]. arXiv:1612.03928, 2016.
[34] Li M, Chen Q, Yan S C. Network in network[J]. arXiv:1312.4400, 2013.
[35] Iandola F N，Han S，Moskewicz M W, et al. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size[J]. arXiv:1602.07360v4, 2016.
[36] Howard A G, Zhu M L, Chen B, et al. MobileNets：efficient convolutional neural networks for mobile vision applications[J]. arXiv:1704.04861, 2017.
[37] Sandler M, Haward A, Zhu M L, et al. MobileNetV2:inverted residuals and linear bottlenecks[C]//Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, Jun 18-23, 2018. Washing-ton: IEEE Computer Society, 2018: 4510-4520.
[38] Zhang X Y, Zhou X Y, Lin M X, et al. ShuffleNet: an extre-mely efficient convolutional neural network for mobile devices[C]//Proceedings of the 2018 IEEE/CVF Conference on Com-puter Vision and Pattern Recognition, Salt Lake City, Jun 18-23, 2018. Washington: IEEE Computer Society, 2018: 6848-6856.
[39] Ma N N, Zhang X Y, Zheng H T, et al. ShuffleNet V2: practical guidelines for efficient CNN architecture design[C]//LNCS 11218: Proceedings of the 15th European Conference on Com-puter Vision, Munich, Sep 8-14, 2018. Cham: Springer Inter-national Publishing, 2018: 122-138.
[40] Mehta S, Rastegari M, Shapiro L, et al. ESPNetv2: a light-weight, power efficient, and general purpose convolutional neural network[J]. arXiv:1811.11431, 2018.
[41] Metha S, Rastegari M, Caspi A, et al. Espnet: efficient spatial pyramid of dilated convolutions for semantic segmentation[C]//LNCS 11214: Proceedings of the 15th European Con-ference on Computer Vision, Munich, Sep 8-14, 2018. Cham: Springer International Publishing, 2018: 561-580.
[42] Wu B C, Wan A, Yue X Y, et al. Shift: a zero FLOP, zero para-meter alternative to spatial convolutions[J]. arXiv:1711.08141, 2017.
[43] Jeon Y H, Kim J M. Constructing fast network through deconstruction of convolution[J]. arXiv:1860.07370, 2018.
[44] Chen W J, Xie D, Zhang Y, et al. You need is a few shifts: designing efficient convolutional neural networks for image classification[J]. arXiv:1903.05285, 2019.
[45] Yu F, Koltun V, Funkhouser T. Dilated residual networks[C]//Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, Jul 21-26, 2017. Washington: IEEE Computer Society, 2017: 636-644.
[46] He K M, Zhang X Y, Ren S Q. Deep residual learning for image recognition[C]//Proceedings of the 2016 IEEE Con-ference on Computer Vision and Pattern Recognition, Las Vegas, Jun 27-30, 2016. Washington: IEEE Computer Society, 2016: 770-778.
[47] Ioffe S, Szegedy C. Batch normalization: accelerating deep network training by reducing internal covariate shift[J]. arXiv:1502.03167, 2015.
[48] Srivastava N, Hinton G, Krizhevsky A, et al. Dropout: a simple way to prevent neural networks from overfitting[J]. The Journal of Machine Learning Research, 2014, 15(1): 1929-1958.
[49] Wang P Q, Chen P F, Yuan Y, et al. Understanding convolu-tion for semantic segmentation[C]//Proceedings of the 2018 IEEE Winter Conference on Applications of Computer Vision, Lake Tahoe, Mar 12-15, 2018. Washington: IEEE Computer Society, 2018: 1451-1460.
[50] Hamaguchi R, Fujita A, Nemoto K, et al. Effective use of dilated convolutions for segmenting small object instances in remote sensing imagery[J]. arXiv:1709.00179, 2017.
[51] He K M, Zhang X Y, Ren S Q, et al. Delving deep into recti-fiers: surpassing human-level performance on ImageNet classi-fication[C]//Proceedings of the 2015 IEEE International Conference on Computer Vision, Santiago, Dec 7-13, 2015. Washington: IEEE Computer Society, 2015: 1026-1034.
[52] Hu J, Shen L, Albanie S, et al. Sqyeeze-and-excitation net-works[J]. arXiv:1709.01507, 2017.
[53] Cai H, Zhu L G, Han S. ProxylessNAS: direct neural architec-ture search on target task and hardware[J]. arXiv:1812.00332， 2018.
[54] Tan M X, Chen B, Pang R M, et al. MnasNet: platform-aware neural architecture search for mobile[C]//Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, Jun 15-20, 2019. Washing-ton: IEEE Computer Society, 2019: 2815-2823.
[55] Wu B C, Dai X L, Zhang P Z, et al. FBNet: hardware-aware efficient ConvNet design via differentiable neural architecture search[J]. arXiv:1812.03443, 2018.
[56] Elsken T, Metzen J H, Hutter F. Neural architecture search: a survey[J]. arXiv:1808.05377, 2018.