Multi-level Dense Capsule Networks

Sai Samarth R. Phaye, Apoorva Sikka, Abhinav Dhall, Deepti R. Bathula

Research output: Chapter in Book/Report/Conference proceedingConference PaperResearchpeer-review

2 Citations (Scopus)

Abstract

Past few years have witnessed an exponential growth of interest in deep learning methodologies with rapidly improving accuracy and reduced computational complexity. In particular, architectures using Convolutional Neural Networks (CNNs) have produced state-of-the-art performances for image classification and object recognition tasks. Recently, Capsule Networks (CapsNets) achieved a significant increase in performance by addressing an inherent limitation of CNNs in encoding pose and deformation. Inspired by such an advancement, we propose Multi-level Dense Capsule Networks (multi-level DCNets). The proposed framework customizes CapsNet by adding multi-level capsules and replacing the standard convolutional layers with densely connected convolutions. A single-level DCNet essentially adds a deeper convolution network, which leads to learning of discriminative feature maps learned by different layers to form the primary capsules. Additionally, multi-level capsule networks uses a hierarchical architecture to learn new capsules from former capsules that represent spatial information in a fine-to-coarser manner, which makes it more efficient for learning complex data. Experiments on image classification task using benchmark datasets demonstrate the efficacy of the proposed architectures. DCNet achieves state-of-the-art performance (99.75%) on the MNIST dataset with approximately twenty-fold decrease in total training iterations, over the conventional CapsNet. Furthermore, multi-level DCNet performs better than CapsNet on SVHN dataset (96.90%), and outperforms the ensemble of seven CapsNet models on CIFAR-10 by +0.31% with seven-fold decrease in the number of parameters. Source codes, models and figures are available at https://github.com/ssrp/Multi-level-DCNet.

Original languageEnglish
Title of host publicationComputer Vision – ACCV 2018
Subtitle of host publication14th Asian Conference on Computer Vision Perth, Australia, December 2–6, 2018 Revised Selected Papers, Part V
EditorsC.V. Jawahar, Hongdong Li, Greg Mori, Konrad Schindler
Place of PublicationCham Switzerland
PublisherSpringer
Pages577-592
Number of pages16
ISBN (Electronic)9783030208738
ISBN (Print)9783030208721
DOIs
Publication statusPublished - 2019
Externally publishedYes
EventAsian Conference on Computer Vision 2018 - Perth, Australia
Duration: 2 Dec 20186 Dec 2018
Conference number: ACCV 2018
http://accv2018.net/

Publication series

NameLecture Notes in Computer Science
PublisherSpringer
Volume11365
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

ConferenceAsian Conference on Computer Vision 2018
Abbreviated title14th
CountryAustralia
CityPerth
Period2/12/186/12/18
Internet address

Keywords

  • Computer vision
  • Deep learning
  • Image recognition

Cite this

Phaye, S. S. R., Sikka, A., Dhall, A., & Bathula, D. R. (2019). Multi-level Dense Capsule Networks. In C. V. Jawahar, H. Li, G. Mori, & K. Schindler (Eds.), Computer Vision – ACCV 2018: 14th Asian Conference on Computer Vision Perth, Australia, December 2–6, 2018 Revised Selected Papers, Part V (pp. 577-592). (Lecture Notes in Computer Science ; Vol. 11365 ). Springer. https://doi.org/10.1007/978-3-030-20873-8_37