Deep Capsule Network Based on Pruning Optimization

被引：0
|
作者：
Zheng X.-P. [1 ]
Liang X. [1 ]
机构：
[1] School of Information, Renmin University of China, Beijing
来源：
Jisuanji Xuebao/Chinese Journal of Computers | 2022年 / 45卷 / 07期
基金：
中国国家自然科学基金;
关键词：
Capsule network; CNN; Deep network; Pruning;
D O I：
10.11897/SP.J.1016.2022.01557
中图分类号：
学科分类号：
摘要：
Convolutional neural network (CNN) has developed rapidly in computer vision tasks, especially in image classification, retrieval, target detection and other fields. CNN can directly preprocess the original pixels, get its effective feature representation and recognize the rules of computer vision. However, CNN adopts a pool strategy, which has a deep understanding of the local characteristics of the object, but ignores the overall spatial structure of the object. A capsule network is a new neural network architecture, which attaches importance to the spatial structure of objects. Capsule network can represent various information of objects by superimposing features into vectors, and has good ability of fitting features. The network architecture can not only recognize the local features of objects, but also retain the hierarchical relationship of the overall spatial features. However, capsule network is not good at complex data. In order to improve the accuracy of image data recognition in capsule network, this paper proposes a deep capsule network based on pruning optimization, which effectively deals with the connection between the capsule network layers, not only reduces the redundant capsule layer, but also improves the generalization ability of the network. The pruning strategy of the new capsule network is to select the low-level capsule strictly, so that only the important and active low-level capsules are reformed into the new capsule layer, and the high-rise capsules are predicted by weighted connection. For image data, the acquired features are insufficient, which makes it difficult to further process in capsule. Therefore, a novel capsule block layer is designed in this paper. The local capsule subset is transformed into a more abstract capsule by 3D convolution core, and the feature representation of more comprehensive data is obtained to construct a more level high-level network. At the same time, the jump connection between the low-level capsule and the high-level capsule can reduce the gradient disappearance problem in the training of the deep model. In addition, the performance test on multiple data sets has achieved good experimental results. © 2022, Science Press. All right reserved.
引用
收藏
页码：1557 / 1570
页数：13
相关论文
共 30 条
  • [1] He K, Zhang X, Ren S, Et al., Spatial pyramid pooling in deep convolutional networks for visual recognition, IEEE Transactions on Pattern Analysis and Machine Intelligence, 37, 9, pp. 1904-1916, (2015)
  • [2] Shin H, Et al., Deep convolutional neural networks for computer-aided detection: CNN architectures, dataset characteristics and transfer learning, IEEE Transactions on Medical Imaging, 35, 5, pp. 1285-1298, (2016)
  • [3] Zhang K, Zuo W, Chen Y, Et al., Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising, IEEE Transactions on Image Processing, 26, 7, pp. 3142-3155, (2017)
  • [4] Chen Y, Jiang H, Li C, Et al., Deep feature extraction and classification of hyperspectral images based on convolutional neural networks, IEEE Transactions on Geoscience and Remote Sensing, 54, 10, pp. 6232-6251, (2016)
  • [5] Ma Y, Liang X, Sheng G, Et al., Noniterative sparse LS-SVM based on globally representative point selection, IEEE Transactions on Neural Networks and Learning Systems, 32, 2, pp. 788-798, (2021)
  • [6] Sabour S, Frosst N, Hinton G E., Dynamic routing between capsules, Advances in Neural Information Processing Systems, pp. 3859-3869, (2017)
  • [7] Rajasegaran J, Jayasundara V, Jayasekara S, Et al., Deepcaps: Going deeper with capsule networks, CoRR
  • [8] Zhao Z, Kleinhans A, Sandhu G, Et al., Capsule networks with max-min normalization, CoRR, (2019)
  • [9] He K, Zhang X, Ren S, Et al., Deep residual learning for image recognition, Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition(CVPR), pp. 770-778, (2016)
  • [10] Dong C, Loy C C, He K, Et al., Image super-resolution using deep convolutional networks, IEEE Transactions on Pattern Analysis and Machine Intelligence, 38, 2, pp. 295-307, (2016)
123