吉林大学学报(工学版) ›› 2022, Vol. 52 ›› Issue (6): 1413-1421.doi: 10.13229/j.cnki.jdxbgxb20210027

• 计算机科学与技术 • 上一篇    

简化型残差结构和快速深度残差网络

杨怀江1,2(),王二帅1,3,隋永新1,2,闫丰1,2,周跃1,2   

  1. 1.中国科学院 长春光学精密机械与物理研究所,长春 130033
    2.长春国科精密光学技术有限公司,长春 130033
    3.中国科学院大学 大珩学院,北京 100049
  • 收稿日期:2021-01-13 出版日期:2022-06-01 发布日期:2022-06-02
  • 作者简介:杨怀江(1966-),男,研究员,博士.研究方向:光学信息融合,深紫外光刻技术.E-mail:994301018@qq.com
  • 基金资助:
    国家重点研发计划项目(2018YFF01011900)

Simplified residual structure and fast deep residual networks

Huai-jiang YANG1,2(),Er-shuai WANG1,3,Yong-xin SUI1,2,Feng YAN1,2,Yue ZHOU1,2   

  1. 1.Changchun Institute of Optics,Fine Mechanics and Physics,Chinese Academy of Sciences,Changchun 130033,China
    2.Changchun National Extreme Precision Optics Co. ,Ltd. ,Changchun 130033,China
    3.Daheng College,University of Chinese Academy of Sciences,Beijing 100049,China
  • Received:2021-01-13 Online:2022-06-01 Published:2022-06-02

摘要:

为解决当前深度残差网络模型训练缓慢的问题,设计了一种新型的残差结构。与典型的残差结构相比,该结构仅含有一个Batch Normalization和ReLU模块,通过减少网络训练过程的计算量降低了耗时,提升了模型训练速度。在常用的CIFAR10/100图像分类数据库上进行了对比实验分析,以该方法构建的深度为110层的网络CIFAR10分类错误率为5.29%,CIFAR100分类错误率为24.80%,典型的110层深度残差网络分类错误率分别为5.75%和26.02%;在训练耗时方面,该方法平均周期耗时为133.47 s,典型的残差网络平均周期耗时为208.26 s,提升了35.91%;结果表明,该网络结构在保证分类性能的基础上极大地提升了训练速度,具有较好的实用价值。

关键词: 图像处理, 图像识别, 图像分类, 卷积神经网络, 深度残差网络

Abstract:

In order to address the problem of slow training of the current deep ResNets model, a novel residual structure is designed. Compared with the typical residual structure, the structure only contains a Batch Normalization and ReLU, which reduces training time and improves the training speed by reducing the amount of calculation in the network training process. The comparative experiments are carried out on the CIFAR10/100 image classification database. The classification error rate of 110 layers networks constructed by this method on CIFAR10 and CIFAR100 is 5.29% and 24.80%, respectively. The classification error rate of 110-ResNet is 5.75% and 26.02%, respectively. Training the network takes 133.47 (this method) and 208.26 (ResNet) seconds per epoch, increased by 35.91%. The results show that the network structure greatly improves the training speed while ensuring the classification performance, and has better practical value.

Key words: image processing, image recognition, image classification, convolutional neural networks, deep residual networks

中图分类号: 

  • TP183

图1

基本型与瓶颈型残差结构"

图2

ReLU示意图"

图3

两种简化型残差结构"

图4

第2类简化型残差结构退化模型"

图5

110层快速深度残差网络模型示意图"

表1

三种残差结构的分类错误率和周期耗时对比"

残差网络分类错误率/%周期耗时/s
基本型7.62±0.1530.07
第1类残差结构7.68±0.2022.08
第2类残差结构10.04±0.1822.10

图6

三种残差结构在CIFAR10上的测试分类错误率曲线"

表2

基本型深度残差网络与快速深度残差网络对比"

残差网络CIFAR10错误率/%CIFAR100错误率/%周期耗时/s
56层ResNet6.3127.9384.22
56层FastResNet5.8126.9158.84
110层ResNet5.7526.02208.26
110层FastResNet5.2924.80133.47

图7

110层ResNet与FastResNet在CIFAR10上的分类错误率曲线"

图8

110层ResNet与FastResNet在CIFAR100上的分类错误率曲线"

1 He Kai-ming, Zhang Xiang-yu, Ren Shao-qing, et al. Deep residual learning for image recognition[C]∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, LasVegas, NV, USA, 2016: 770-778.
2 He Kai-ming, Zhang Xiang-yu, Ren Shao-qing, et al. Identity mappings in deep residual networks[C]∥European Conference on Computer Vision, Cham, 2016: 630-645.
3 Targ S, Almeid A D, Lyman K. Resnet in resnet: Generalizing residual architectures[EB/OL]. [2016-03-25].
4 Zagoruyko S, Komodakis N. Wide residual networks[EB/OL]. [2017-06-14].
5 Veit A, Wilber M, Belongie S. Residual networks behave like ensembles of relatively shallow networks[C]∥Proceedings of the 30th International Conference on Neural Information Processing Systems. Newyork: Curran Associates Inc, 2016:550-558.
6 Wu Z F, Shen C H, van den Hengel A. Wider or deeper: revisiting the resnet model for visual recognition[EB/OL]. [2016-11-30].
7 王若瑜. 基于Resnet-50的智能驾驶红绿灯分类研究[J]. 电子测试,2019, 25():143-145.
Wang Ruo-yu. Research on classification of intelligent driving traffic light based on Resnet-50[J]. Electronic Test, 2019, 25(Sup.1):143-145.
8 宋倩, 黄昶, 余慧瑶. 基于TensorFlow的交通标志形状识别[J]. 信息通信, 2017, 32(12): 286-288.
Song Qian, Huang Chang, Yu Hui-yao. Traffic sign shape recognition based on tensorflow[J]. Information & Communications, 2017, 32(12): 286-288.
9 王文成, 蒋慧, 乔倩, 等. 基于ResNet50网络的十种鱼类图像分类识别研究[J]. 农村经济与科技, 2019, 30(19): 60-62.
Wang Wen-cheng, Jiang Hui, Qiao Qian, et al. Research on classification and recognition of ten fish images based on Resnet-50[J]. Rural Economy and Science-Technology, 2019, 30(19): 60-62.
10 常川. 基于ResNet深度网络的人类蛋白质图谱图像分类方法研究[J]. 医学信息学, 2019, 40(7): 45-49.
Chang Chuan. Study of image classification method of human protein atlas based on resnet deep network[J]. Journal of Medical Intelligence, 2019, 40(7): 45-49.
11 王恒,李霞,刘晓芳, 等. 基于ResNet50网络的乳腺癌病理图像分类研究[J]. 中国计量大学学报, 2019, 30(1): 72-77.
Wang Heng, Li Xia, Liu Xiao-fang, et al. Classification of breast cancer histopathological images based on ResNet50 network[J]. Journal of China Jiliang University, 2019, 30(1): 72-77.
12 张枫, 田联房, 杜启亮. 基于残差网络与中心损失的人脸识别[J].计算机工程与设计, 2019, 40(6): 1689-1695.
Zhang Feng, Tian Lian-fang, Du Qi-liang. Face recognition based on Resnet and center loss[J]. Computer Engineering and Design, 2019, 40(6): 1689-1695.
13 刘媛. 基于深度学习的说话人识别研究[D]. 上海: 上海交通大学信息与通信工程学院,2015.
Liu Yuan. Research on deep leaning based speaker recognition[D]. Shanghai: School of Electronic Information and Electrical Engineering, Shanghai Jiao Tong University, 2015.
14 朱超平, 杨艺. 基于YOLO2和ResNet算法的监控视频中的人脸检测与识别[J]. 重庆理工大学学报: 自然科学, 2018, 32(8): 170-175.
Zhu Chao-ping, Yang Yi. Face Detection and recognition in monitoring video based on YOLO2 and ResNet algorithm[J]. Journal of Chongqing University of Technology(Natural Science), 2018, 32(8): 170-175.
15 熊才华,巩言丽,廉华, 等. 基于ResNet-50改进的Faster R-CNN手势识别算法[J]. 计算机时代, 2019, 36(9): 1-4.
Xiong Cai-hua, Gong Yan-li, Lian Hua, et al. An improved Faster R-CNN hand gesture recognition algorithm based on ResNet-50[J]. Computer Era, 2019, 36(9):1-4.
16 刘子毅. 基于图谱特征分析的农业虫害检测方法研究[D]. 杭州: 浙江大学生物系统工程与食品科学学院, 2017.
Liu Zi-yi. Detection of agricultural pest insects based on imaging and spectral feature analysis[D]. Hangzhou: College of Biosystems Engineering and Food Science, Zhejiang University, 2017.
17 张银松,赵银娣,袁慕策. 基于改进Faster-RCNN模型的粘虫板图像昆虫识别与计数[J]. 中国农业大学学报, 2019, 24(5): 115-122.
Zhang Yin-song, Zhao Yin-di, Yuan Mu-ce. Insect identification and counting based on an improved Faster-RCNN model of the sticky board image[J]. Journal of China Agricultural University, 2019, 24(5): 115-122.
18 Huang Gao, Sun Yu, Liu Zhuang, et al. Deep networks with stochastic depth[C]∥European Conference on Computer Vision, Cham, 2016: 646-661.
19 Glorot X, Bengio Y. Understanding the difficulty of training deep feedforward neural networks[C]∥Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, Canada, 2010:249-256.
20 Krizhevsky A, Sutskever I, Hinton G. Imagenet classification with deep convolutional neural networks[J]. Communications of the ACM, 2017, 60(6): 84-90.
21 Lin Min, Chen Qiang, Yan Shui-cheng. Network in network[EB/OL]. [2014-03-04].
22 Springenberg J, Dosovitskiy A, Brox T, et al. Striving for simplicity: the all convolutional net[EB/OL]. [2015-04-13].
23 Szegedy C, Liu W, Jia Y Q, et al. Going deeper with convolutions[C]∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Boston, MA, 2015: 1-9.
24 Ioffe S, Szegedy C. Batch normalization: accelerating deep network training by reducing internal covariate shift[EB/OL]. [2015-03-02].
25 Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition[EB/OL]. [2015-04-10].
26 Santurkar S, Tsipras D, Ilyas A, et al. How does batch normalization help optimization?[EB/OL]. [2021-01-02].
[1] 高明华,杨璨. 基于改进卷积神经网络的交通目标检测方法[J]. 吉林大学学报(工学版), 2022, 52(6): 1353-1361.
[2] 贾海洋,夏睿,吕安祺,管噌璇,陈娟,王雷. 基于模板融合的超声医学影像全景拼接方法[J]. 吉林大学学报(工学版), 2022, 52(4): 916-924.
[3] 李先通,全威,王华,孙鹏程,安鹏进,满永兴. 基于时空特征深度学习模型的路径行程时间预测[J]. 吉林大学学报(工学版), 2022, 52(3): 557-563.
[4] 李向军,涂洁莹,赵志宾. 基于多尺度融合卷积神经网络的熔解曲线有效性分类[J]. 吉林大学学报(工学版), 2022, 52(3): 633-639.
[5] 王学智,李清亮,李文辉. 融合迁移学习的土壤湿度预测时空模型[J]. 吉林大学学报(工学版), 2022, 52(3): 675-683.
[6] 张龙,徐天鹏,王朝兵,易剑昱,甄灿壮. 基于卷积门控循环网络的齿轮箱故障诊断[J]. 吉林大学学报(工学版), 2022, 52(2): 368-376.
[7] 李茂月,刘硕,田帅,肖桂风. 薄壁件铣削加工颤振的图像特征提取与识别[J]. 吉林大学学报(工学版), 2022, 52(2): 425-432.
[8] 高文志,王彦军,王欣伟,张攀,李勇,董阳. 基于卷积神经网络的柴油机失火故障实时诊断[J]. 吉林大学学报(工学版), 2022, 52(2): 417-424.
[9] 段亮,宋春元,刘超,魏苇,吕成吉. 基于机器学习的高速列车轴承温度状态识别[J]. 吉林大学学报(工学版), 2022, 52(1): 53-62.
[10] 钟辉,康恒,吕颖达,李振建,李红,欧阳若川. 基于注意力卷积神经网络的图像篡改定位算法[J]. 吉林大学学报(工学版), 2021, 51(5): 1838-1844.
[11] 王德兴,吴若有,袁红春,宫鹏,王越. 基于多尺度注意力融合和卷积神经网络的水下图像恢复[J]. 吉林大学学报(工学版), 2021, 51(4): 1396-1404.
[12] 王康,姚猛,李立犇,李建桥,邓湘金,邹猛,薛龙. 基于月面表取采样触月压痕的月壤力学状态分析[J]. 吉林大学学报(工学版), 2021, 51(3): 1146-1152.
[13] 许骞艺,秦贵和,孙铭会,孟诚训. 基于改进的ResNeSt驾驶员头部状态分类算法[J]. 吉林大学学报(工学版), 2021, 51(2): 704-711.
[14] 李健,刘孔宇,任宪盛,熊琦,窦雪峰. 基于自适应阈值的Canny算法在MRI边缘检测中的应用[J]. 吉林大学学报(工学版), 2021, 51(2): 712-719.
[15] 刘富,刘璐,侯涛,刘云. 基于优化MSR的夜间道路图像增强方法[J]. 吉林大学学报(工学版), 2021, 51(1): 323-330.
Viewed
Full text


Abstract

Cited

  Shared   
  Discussed   
No Suggested Reading articles found!