融合注意力和胶囊池化的轻量型胶囊网络

doi:10.16180/j.cnki.issn1007-7820.2024.05.001

摘要/Abstract

摘要：

针对胶囊网络特征信息传播低效性和路由过程存在较大计算开销等问题,文中提出了一种融合注意力和胶囊池化的轻量型胶囊网络。该网络主要有以下两方面的优势:1)提出了胶囊注意力。将注意力作用于初级胶囊层,增强对重要胶囊的关注,提高低级胶囊对高级胶囊预测的准确性;2) 提出新的胶囊池化。在初级胶囊层所有特征图的对应位置筛选出权重最大的胶囊,在减少模型参数量的同时以少量的重要胶囊表示有效特征信息。公共数据集的结果表明,提出的胶囊网络在CIFAR10上达到92.60%的精度,并在复杂数据集上具有良好的白盒对抗攻击鲁棒性。此外,提出的胶囊网络在AffNIST数据集上达到95.74%的精度,具有较好的仿射变换鲁棒性。计算效率结果表明,所提网络的浮点运算量比传统胶囊网络减少了31.3%,参数量减少了41.9%。

关键词: 深度学习, 图像分类, 胶囊网络, 胶囊池化, 注意力机制, 鲁棒性, 对抗攻击, 轻量型

Abstract:

In view of the inefficiency of feature information propagation in capsule networks and the huge computational overhead in the routing process, a graph pooling capsule network that combines attention and capsule pooling is proposed. The network mainly has the following two advantages: 1) The capsule attention is proposed, and the attention is applied to the primary capsule layer, which enhances the attention to the important capsules, and improves the accuracy of the prediction of the lower capsules to the higher capsules; 2) A new capsule pooling is proposed. The capsule with the largest weight is screened out at the corresponding positions of all feature maps in the primary capsule layer, and the effective feature information is represented by a small number of important capsules while reducing the number of model parameters. Results on public data sets show that the proposed capsule network achieves the accuracy of 92.60% on CIFAR10 and has excellent robustness against white-box adversarial attacks on complex datasets. In addition, the proposed capsule network achieves 95.74% accuracy on the AffNIST data set with superior affine transformation robustness. The calculation efficiency results show that the amount of floating-point operations of the proposed capsule is reduced by 31.3% and the number of parameters is reduced by 41.9% when compared with traditional CapsNet.

Key words: deep learning, image classification, capsule network, capsule pooling, attention mechanism, robustness, adversarial attack, lightweight

中图分类号:

TP391.4

朱子豪, 宋燕. 融合注意力和胶囊池化的轻量型胶囊网络[J]. 电子科技, 2024, 37(5): 1-8.

ZHU Zihao, SONG Yan. Lightweight Capsule Network Fusing Attention and Capsule Pooling[J]. Electronic Science and Technology, 2024, 37(5): 1-8.

图/表 10

图1

表1

表2

表3

表4

图2

图3

图4

图5

表5

参考文献 21

[1]	Lecun Y, Bottou L, Bengio Y, et al. Gradient-based learning applied to document recognition[J]. Proceedings of the IEEE, 1998, 86(11): 2278-232.
[2]	刘林嵩, 仝明磊, 吴东亮. SA-CapsNet:自注意力胶囊网络[J]. 计算机应用研究, 2021, 38(10):3005-3008,3039.
	Liu Linsong, Gong Minglei, Wu Dongliang. SA-CapsNet: Self-attention capsule network[J]. Application Research of Computers, 2021, 38(10):3005-3008,3039.
[3]	张伟, 刘娜, 江洋, 等. 基于YOLO神经网络的垃圾检测与分类[J]. 电子科技, 2022, 35(10):45-50.
	Zhang Wei, Liu Na, Jiang Yang, et al. Garbage detection and classification based on YOLO neural network[J]. Electronic Science and Technology, 2022, 35(10):45-50.
[4]	房巾莉, 吕毅斌, 王樱子, 等. 基于水平集的医学图像分割算法[J]. 电子科技, 2021, 34(2):12-20.
	Fang Jinli, Lü Yibin, Wang Yingzi, et al. A novel medical image segmentation algorithm based on level set[J]. Electronic Science and Technology, 2021, 34(2):12-20.
[5]	李业良, 张二华, 唐振民. 基于混合式注意力机制的语音识别研究[J]. 计算机应用研究, 2020, 37(1): 131-134.
	Li Yeliang, Zhang Erhua, Tang Zhenmin. Research on speech recognition based on hybrid attention mechanism[J]. Application Research of Computers, 2020, 37(1):131-134.
[6]	Sabour S, Frosst N, Hinton G E. Dynamic routing between capsules[C]. Long Beach: Proceedings of Neural Information Processing Systems, NeurIPS, 2017:3856-3866.
[7]	Xiang C, Zhang L, Tang Y, et al. MS-CapsNet: A novel multiscale capsule network[J]. IEEE Signal Processing Letters, 2018, 25(12):1850-1854.
[8]	Rajasegaran J, Jayasundara V, Jayasekara S, et al. Deepcaps: Going deeper with capsule networks[C]. Los Angeles: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019:10725-10733.
[9]	Gu J D. Interpretable graph capsule networks for object recognition[C]. Vancouver: Proceedings of the AAAI Conference on Artificial Intelligence, 2021:1469-1477.
[10]	Huang R Y, Li J P, Wang S H, et al. A robust weight-shared capsule network for intelligent machinery fault diagnosis[J]. IEEE Transactions on Industrial Informatics, 2020, 16(10):6466-6475.
[11]	Jia X F, Li J Q, Zhao B T, et al. Res-CapsNet:Residual capsule network for data classification[J]. Neural Process Letter, 2022, 54(5):4229-4245.
[12]	He K M, Zhang X Y, Ren S Q, et al. Deep residual learning for image recognition[C]. Seattle: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016:770-778.
[13]	Choi J, Seo H, Im S, et al. Attention routing between capsules[C]. Seoul: IEEE/CVF International Conference on Computer Vision Workshops, 2019:1981-1989.
[14]	Hinton G E, Sabour S, Frosst N. Matrix capsules with EM routing[C]. Vancouver: International Conference on Learning Representations, 2018:1-15.
[15]	Ribeiro F D S, Leontidis G, Kollias S. Capsule routing via variational bayes[C]. New York: AAAI Conference on Artificial Intelligence, 2020:3749-3756.
[16]	Hahn T, Pyeon M, Kim G. Self-routing capsule networks[C]. Vancouver: Advances in Neural Information Processing Systems, 2019:7658-7667.
[17]	Gu J D, Tresp V. Improving the robustness of capsule networks to image affine transformations[C]. Seattle: IEEE Conference on Computer Vision and Pattern Recognition, 2020:7283-7291.
[18]	Zhang S F, Zhao W, Wu X F, et al. Fast dynamic routing based on weighted kernel density estimation[C]. Daegu: International Symposium on Artificial Intelligence and Robotics, 2021: 5281-5297.
[19]	Jeong T, Lee Y, Kim H. Ladder capsule network[C]. New York: International Conference on Machine Learning, 2019:3071-3079.
[20]	Lenssen J E, Fey M, Libuschewski P. Group equivariant capsule networks[C]. Montréal: Advances in Neural Information Processing Systems, 2018:8844-8853.
[21]	Kosiorek A R, Sabour S, Teh Y W, et al. Stacked capsule autoencoders[C]. Vancouver: Advances in Neural Information Processing Systems, 2019:15486-15496.

模型方法	MNIST		FashionMNIST		SVHN		CIFAR10		SmallNorb
模型方法	错误率/%	参数量/MB	错误率/%	参数量/MB	错误率/%	参数量/MB	错误率/%	参数量/MB	错误率/%	参数量/MB
Attn-Routing^[13]	0.54	5.31	-	-	-	-	12.71	9.600	-	-
EM-Routing^[14]	-	-	-	-	-	-	11.90	0.450	1.80	0.30
VB-Routing^[15]	-	-	5.20	0.17	3.90	0.320	11.20	0.320	1.60	0.17
Self-routing^[16]	-	-	-	-	3.12	140.300	7.86	140.300	-	-
DRCaps^[6]	0.32	6.10	5.45	6.10	3.85	6.100	9.94	6.100	3.52	6.10
Aff-CapsNet^[17]	0.46	8.20	7.47	8.20	7.85	8.200	23.72	8.200	-	-
MS-CapsNet^[7]	-	-	7.30	10.80	-	-	24.30	11.200	-	-
FRMS^[18]	0.42	1.2	6.00	1.20			15.60	1.200
GraCapsnets^[9]	-	-	-	-	2.98	2.475	7.99	2.475	-	-
CAPCaps(本文)	0.29	3.00	5.41	3.00	3.46	3.540	7.40	3.540	4.90	3.54

模型	CIFAR10/%	SVHN/%
Baseline	9.20	4.09
Baseline+(CA)	8.86	3.70
Baseline+(CP)	8.90	3.77
Baseline+(CA+CP)	7.40	3.46

模型方法	MNIST分类精度/%	AffNIST分类精度/%
CNN^[14]	99.20	85.90
Dynamic^[6]	99.20	79.00
LadderCaps^[19]	99.30	87.80
GCaps^[20]	98.40	89.10
Attn-Routing^[13]	99.50	91.60
SCAE^[21]	98.50	92.20
Aff-CapsNet^[17]	99.23	93.12
EM-Routing^[14]	99.20	93.10
DRCaps	99.20	95.42
CAPCaps	99.20	93.25
CAPCaps	99.56	95.74

模型方法	MNIST测试精度/%	Rotate-MNIST测试精度/%
AvgPool	99.48	95.73
Conv	99.47	96.49
DRCaps	99.50	97.25
CAPCaps	99.49	97.92

模型方法	Params/MB	Flops/MFlops
DRCaps	6.100	552.38
GraCapsnets	2.475	380.00
CAPCaps	3.540	380.00