计算机应用 ›› 2020, Vol. 40 ›› Issue (12): 3451-3457.DOI: 10.11772/j.issn.1001-9081.2020060882

• 2020年中国粒计算与知识发现学术会议(CGCKD 2020) • 上一篇    下一篇

基于跨通道交叉融合和跨模块连接的轻量级卷积神经网络

陈力, 丁世飞, 于文家   

  1. 中国矿业大学 计算机科学与技术学院, 江苏 徐州 221116
  • 收稿日期:2020-06-19 修回日期:2020-08-24 出版日期:2020-12-10 发布日期:2020-10-20
  • 通讯作者: 丁世飞(1963-),男,山东青岛人,教授,博士,CCF会员,主要研究方向:人工智能、机器学习、深度强化学习。dingsf@cumt.edu.cn
  • 作者简介:陈力(1993-),男,山东邹城人,硕士研究生,主要研究方向:深度学习、图像处理;于文家(1994-),男,辽宁本溪人,硕士研究生,主要研究方向:深度学习
  • 基金资助:
    国家自然科学基金资助项目(61672522,61976216,61379101)。

Lightweight convolutional neural network based on cross-channel fusion and cross-module connection

CHEN Li, DING Shifei, YU Wenjia   

  1. School of Computer Science and Technology, China University of Mining and Technology, Xuzhou Jiangsu 221116, China
  • Received:2020-06-19 Revised:2020-08-24 Online:2020-12-10 Published:2020-10-20
  • Supported by:
    This work is partially supported by the National Natural Science Foundation of China (61672522, 61976216, 61379101).

摘要: 针对传统卷积神经网络参数量过多、计算复杂度高的问题,提出了基于跨通道交叉融合和跨模块连接的轻量级卷积神经网络架构C-Net。首先,提出了跨通道交叉融合的方法,它在一定程度上克服了分组卷积中各分组之间存在缺乏信息流动的问题,简单高效地实现了不同分组之间的信息通信;其次,提出了一种跨模块连接的方法,它克服了传统轻量级架构中各基本构建块之间彼此独立的缺点,实现了同一阶段内具有相同分辨率特征映射的不同模块之间的信息融合,从而增强了特征提取能力;最后,基于提出的两种方法设计了一种新型的轻量级卷积神经网络架构C-Net。C-Net在Food_101数据集上的准确率为69.41%,在Caltech_256数据集上的准确率为63.93%。实验结果表明,与目前先进的轻量级卷积神经网络模型相比,C-Net降低了存储开销和计算复杂度。在Cifar_10数据集上的消融实验验证了所提出的两种方法的有效性。

关键词: 卷积神经网络, 轻量级, 分组卷积, 跨通道交叉融合, 快捷连接, 跨模块连接

Abstract: In order to solve the problems of too many parameters and high computational complexity of traditional convolutional neural networks, a lightweight convolutional neural network architecture named C-Net based on cross-channel fusion and cross-module connection was proposed. Firstly, a method called cross-channel fusion was proposed. With it, the shortcoming of lacking information flow between different groups of grouped convolution was solved to a certain extent, and the information communication between different groups was realized efficiently and easily. Then, a method called cross-module connection was proposed. With it, the shortcoming that the basic building blocks in the traditional lightweight architecture were independent to each other was overcome, and the information fusion between different modules with the same resolution feature mapping within the same stage was achieved, enhancing the feature extraction capability. Finally, a novel lightweight convolutional neural network architecture C-Net was designed based on the two proposed methods. The accuracy of C-Net on the Food_101 dataset is 69.41%, and the accuracy of C-Net on the Caltech_256 dataset is 63.93%. Experimental results show that C-Net reduces the memory cost and computational complexity in comparison with the state-of-the-art lightweight convolutional neural network models. The ablation experiment verifies the effectiveness of the two proposed methods on the Cifar_10 dataset.

Key words: Convolutional Neural Network (CNN), lightweight, grouped convolution, cross-channel fusion, shortcut connection, cross-module connection

中图分类号: