基于Agglomerator特征提取的非迭代路由胶囊网络

doi:10.16180/j.cnki.issn1007-7820.2025.08.004

电子科技 ›› 2025, Vol. 38 ›› Issue (8): 27-32.doi: 10.16180/j.cnki.issn1007-7820.2025.08.004

基于Agglomerator特征提取的非迭代路由胶囊网络

倪庭轩, 宋燕()

上海理工大学光电信息与计算机工程学院,上海 200093

收稿日期:2024-01-15 修回日期:2024-02-08 出版日期:2025-08-15 发布日期:2025-07-10
通讯作者: 宋燕(1979-),女,E-mail:sonya@usst.edu.cn,博士,教授。研究方向:模式识别、数据分析和预测控制等。
作者简介:倪庭轩(1997-),男,硕士研究生。研究方向:图像处理。
基金资助:
国家自然科学基金(62073223);上海市自然科学基金(22ZR1443400)

A Non-Iterative Routing CapsNet Based on Agglomerator Feature Extraction

NI Tingxuan, SONG Yan()

School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China

Received:2024-01-15 Revised:2024-02-08 Online:2025-08-15 Published:2025-07-10
Supported by:
National Natural Science Foundation of China(62073223);Shanghai Natural Science Foundation(22ZR1443400)

摘要/Abstract

摘要：

针对胶囊网络特征提取器可解释性低的问题,文中提出了一种融合DenseCap(Dense Capsule)和Agglomerator的新特征提取器。将密集连接的低级高级特征与Agglomerator的局部整体特征相结合,使相邻两层特征对应局部和整体,提高了可解释性。DenseCap和Agglomerator的并联连接方式使模型结构更紧凑,能够减少可训练参数。将绝对位置编码与Agglomerator密集连接,在计算相对注意力时保留绝对值编码和相对位置编码的优点,保持平移等变性。实验结果表明,相较于胶囊网络以及原始Agglomerator,Agg-CapsNet(Agglomerator Capsule Network)在CIFAR10、MNIST、Fashion-MNIST和SmallNorb方面精度更好。在位置编码的平移实验中,通过可视化证明Agg-CapsNet具有平移等变性。

关键词: 特征提取, 胶囊网络, Agglomerator, 位置编码, 等变性, 注意力, 局部整体特征, 对比学习

Abstract:

In view of the problem of low interpretability of the feature extractor of capsule network, a new feature extractor combining DenseCap(Dense Capsule) and Agglomerator is proposed in this study. By combining the densely connected low-level and high-level features with the local global features of Agglomerator, the adjacent two layers of features correspond to the local and the whole, which improves the interpretability. The parallel connection of DenseCap and Agglomerator makes the model structure more compact and reduces trainable parameters. Dense connection of absolute position coding with Agglomerator preserves the advantages of absolute value coding and relative position coding when calculating relative attention, and maintains translational isotropy. The experimental results show that compared with the Capsule Network and the original Agglomerator, the Agg-CapsNet (Agglomerator Capsule Network) has better accuracy in terms of CIFAR10, MNIST, Fashion MNIST and SmallNorb. In translation experiments of position coding, Agg-CapsNet is proved to have translational isotropy by visualization.

Key words: feature extraction, capsule network, Agglomerator, position coding, invariance, attention, part and whole feature, contrastive learning

中图分类号:

TP183

倪庭轩, 宋燕. 基于Agglomerator特征提取的非迭代路由胶囊网络[J]. 电子科技, 2025, 38(8): 27-32.

NI Tingxuan, SONG Yan. A Non-Iterative Routing CapsNet Based on Agglomerator Feature Extraction[J]. Electronic Science and Technology, 2025, 38(8): 27-32.

图/表 8

图1

图2

图3

表1

表2

表3

图4

图5

参考文献 18

[1]	Sabour S, Frosst N, Hinton G E. Dynamic routing between capsules[C]. Long Beach: Advances in Neural Information Processing Systems, 2017:3856-3866.
[2]	Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition[C]. San Diego: International Conference on Learning Representation, 2015:1-14.
[3]	He K M, Zhang X Y, Ren S Q, et al. Deep residual learning for image recognition[C]. Las Vegas: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016:770-778.
[4]	Huang G, Liu Z, Laurens V D M, et al. Densely connected convolutional networks[C]. Honolulu: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017:4700-4708.
[5]	Zeng R, Song Y. A fast routing capsule network with improved dense blocks[J]. IEEE Transactions on Industrial Informatics, 2021, 18(7):4383-4392.
[6]	Zeng R, Zhu J L, Song Y, et al. Enhanced capsule networks with improved residual modules[C]. Beijing: IEEE International Conference on Unmanned Systems, 2021:39-44.
[7]	Rajasegaran J, Jayasundara V, Jayasekara S, et al. DeepCaps:Going deeper with capsule networks[C]. Long Beach: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019:10725-10733.
[8]	Hinton G E. How to represent part-whole hierarchies in a neural network[J]. Neural Computation, 2023, 35(3):413-452.
[9]	Alexey D, Lucas B, Alexander K, et al. An image is worth 16×16 words:Transformers for image recognition at scale[C]. Online: International Conference on Learning Representations, 2021:1-21.
[10]	Garau N, Bisagno N, Sambugaro Z, et al. Interpretable part-whole hierarchies and conceptual-semantic relationships in neural networks[C]. New Orleans: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022:13689-13698.
[11]	Bromley J, Guyon I, Yang L C, et al. Signature verification using a "Siamese" time delay neural network[C]. Denver: Advances in Neural Information Processing Systems, 1993:737-744.
[12]	Md A I, Sen J, Neil D B B. How much position information do convolutional neural networks encode[C]. Online: International Conference on Learning Representations, 2020:1-11.
[13]	Xu K, Ba J, Kiros R, et al. Show,attend and tell:Neural image caption generation with visual attention[C]. Lille: Proceedings of the Thirty-second International Conference on Machine Learning, 2015:2048-2057.
[14]	Chen T, Kornblith S, Norouzi M, et al. A simple framework for contrastive learning of visual representations[C]. Online: Proceedings of the Thirty-seventh International Conference on Machine Learning, 2020:1597-1607.
[15]	Hinton G E, Sabour S, Frosst N. Matrix capsules with EM routing[C]. Vancouver: International Conference on Learning Representations, 2018:1-15.
[16]	Tankovich V, Hane C, Zhang Y D, et al. HITNet: Hierarchical iterative tile refinement network for real-time stereo matching[C]. Online: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021:14362-14372.
[17]	Hahn T, Pyeon M, Kim G. Self-routing capsule networks[C]. Vancouver: Advances in Neural Information Processing Systems, 2019:7658-7667.
[18]	Su J L, Ahmed M, Lu Y, et al. RoFormer:Enhanced transformer with rotary position embedding[J]. Neurocomputing, 2024, 56(8):127063-127075.

T	L	参数量(浮点数数量)	精度/%
6	32	6.97×10⁶	92.75
6	6	3.55×10⁶	92.62
2	6	3.27×10⁶	92.49
10	6	3.83×10⁶	92.31

模型	参数量 (浮点数数量)	CIFAR10/ %	MNIST/ %	Fashion-MNIST/ %
动态路由^[1]	6.80×10⁶	89.40	99.75	93.60
HitNet^[16]	8.20×10⁶	73.30	99.76	92.30
EM路由^[15]	0.30×10⁶	88.10	99.56	93.86
Agglomerator^[10]	72.00×10⁶	88.95	99.70	92.57
Agg-CapsNet	3.55×10⁶	92.97	99.77	94.24

模型	高度角		水平角
模型	熟悉/%	陌生/%	熟悉/%	陌生/%
动态路由^[1]	88.14	73.17	92.77	82.92
SR^[17]	92.38	80.14	94.04	84.09
FR^[5]	92.01	81.47	95.97	84.44
Agg-CapsNet	92.27	80.48	94.58	85.39

基于Agglomerator特征提取的非迭代路由胶囊网络

A Non-Iterative Routing CapsNet Based on Agglomerator Feature Extraction

RichHTML

PDF (PC)

赞

可视化

摘要/Abstract

引用本文

使用本文

图/表 8

参考文献 18

相关文章 15

Metrics

本文评价

推荐阅读 0

[1]	熊子恒, 张轩雄. 基于深度学习的红外与可见光图像匹配[J]. 电子科技, 2025, 38(9): 1-8.
[2]	龚宇, 吴鹏. 基于交叉融合编码器的Transformer图像特征提取网络[J]. 电子科技, 2025, 38(9): 20-25.
[3]	齐浩东, 程晓颖, 李海生, 陈晓, 徐鑫炯. 基于改进YOLOv7的焊缝缺陷检测算法[J]. 电子科技, 2025, 38(9): 79-84.
[4]	赵开, 胡春燕, 李菲菲. 基于高分辨率网络的轻量化人体姿态估计算法[J]. 电子科技, 2025, 38(9): 93-100.
[5]	程伟豪, 杨晓城, 武林, 阎敬业, 蒋明峰, 魏波. 基于改进Transformer的综合孔径辐射计重构算法[J]. 电子科技, 2025, 38(8): 94-100.
[6]	沈依宁, 王一然, 吴聪. 基于深度学习的关系抽取研究进展[J]. 电子科技, 2025, 38(7): 40-49.
[7]	唐玉良, 张轩雄. 基于深度学习的车位检测方[J]. 电子科技, 2025, 38(6): 23-29.
[8]	陈清江, 杨双. 基于多层特征增强的双分支图像去雾算法[J]. 电子科技, 2025, 38(6): 30-38.
[9]	刘雪莉, 杜洪波, 袁雪丰, 朱立军. 基于CBAM残差块结合纹理采样器的水墨迁移算法[J]. 电子科技, 2025, 38(6): 65-73.
[10]	沈庭铅, 鲁玉军, 辛昊, 吴涵超, 汪仕男. 基于改进YOLOv5的钢板表面缺陷检测[J]. 电子科技, 2025, 38(6): 82-88.
[11]	陈漫漫, 于莲芝. 结合共注意网络的深度BiGRU和DPCS情感分析模型[J]. 电子科技, 2025, 38(5): 22-30.
[12]	胡锦磊, 田恩刚, 瞿枫. 融合通道全局注意力机制的双路径实时语义分割[J]. 电子科技, 2025, 38(5): 83-88.
[13]	王勇, 杨义龙, 范晓晖, 周雷, 孔祥勇. 基于迁移学习和改进EfficientNet-B0的脑肿瘤分类算法[J]. 电子科技, 2025, 38(4): 46-51.
[14]	周凯, 于莲芝. 融合注意力机制的弱监督语义分割自激活方法[J]. 电子科技, 2025, 38(4): 80-86.
[15]	王紫祎, 陈世平. 基于图对比学习的自监督网络流量检测模型[J]. 电子科技, 2025, 38(3): 22-31.