Skip to content

A consistency-aware deep capsule network for hierarchical multi-label image classification

Notifications You must be signed in to change notification settings

tasrif-khondaker/HD-CapsNet

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

69 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

A consistency-aware deep capsule network for hierarchical multi-label image classification

This is the official implementation of the paper titled "A consistency-aware deep capsule network for hierarchical multi-label image classification" by Khondaker Tasrif Noor, Antonio Robles-Kelly, Leo Yu Zhang, Mohamed Reda Bouadjenek, and Wei Luo. The paper is available on Neurocomputing Journal.

Abstract

Hierarchical classification is a significant challenge in computer vision due to the logical order and interconnectedness of multiple labels. This paper presents HD-CapsNet, a novel neural network architecture based on deep capsule networks, specifically designed for hierarchical multi-label classification(HMC). By incorporating a tree-like hierarchical structure, HD-CapsNet is designed to leverage the inherent ontological order within the hierarchical label tree, thereby ensuring classification consistency across different levels. Additionally, we introduce a specialized loss function that promotes accurate hierarchical relationships while penalizing inconsistencies. This not only enhances classification performance but also strengthens the network’s robustness. We rigorously evaluate HD-CapsNet’s efficacy by benchmarking it against existing HMC methods across six diverse datasets: Fashion-MNIST, Marine-Tree, CIFAR-10, CIFAR-100, Caltech-UCSD Birds-200-2011, and Stanford Cars. Our results conclusively demonstrate that HD-CapsNet excels in learning hierarchical relationships and significantly outperforms the competition in various image classification tasks.

Results:


Dataset Models Total Trainable params (M) Accuracy Coarse Accuracy Medium Accuracy Fine Hierarchical Precision Hierarchical Recall Hierarchical F1-Score Consistency Exact Match
Fashion MNIST HD-CapsNet 4.82 99.92% 97.79% 94.83% 97.51% 97.54% 97.52% 99.84% 94.74%
Fashion MNIST HD-CapsNet $~\dagger$ 4.82 99.89% 97.78% 94.92% 97.53% 97.59% 97.55% 99.70% 94.77%
Fashion MNIST HD-CapsNet $~\ddagger$ 4.73 99.91% 97.63% 94.66% 97.40% 97.42% 97.41% 99.87% 94.60%
CIFAR-10 HD-CapsNet 5.23 98.79% 94.28% 91.22% 94.74% 94.89% 94.80% 99.18% 90.95%
CIFAR-10 HD-CapsNet $~\dagger$ 5.23 98.71% 94.01% 90.97% 94.53% 94.73% 94.62% 98.99% 90.58%
CIFAR-10 HD-CapsNet $~\ddagger$ 4.84 98.76% 93.36% 90.26% 94.09% 94.30% 94.18% 98.94% 89.85%
CIFAR-100 HD-CapsNet 7.85 86.93% 79.31% 66.38% 77.43% 79.20% 78.12% 89.80% 63.80%
CIFAR-100 HD-CapsNet $~\dagger$ 7.85 86.81% 78.73% 66.23% 77.10% 79.02% 77.85% 88.62% 63.36%
CIFAR-100 HD-CapsNet $~\ddagger$ 5.55 86.57% 78.33% 57.08% 73.86% 75.00% 74.31% 92.51% 56.10%
Marine Tree HD-CapsNet 13.58 89.88% 78.60% 57.15% 75.02% 76.04% 75.44% 94.47% 55.59%
Marine Tree HD-CapsNet $~\dagger$ 13.58 89.50% 77.57% 53.75% 73.29% 74.76% 73.88% 92.37% 51.85%
Marine Tree HD-CapsNet $~\ddagger$ 5.97 86.98% 77.82% 55.04% 73.35% 75.76% 74.36% 86.95% 49.34%
CU Bird HD-CapsNet 106.01 40.42% 21.61% 13.39% 23.47% 30.33% 26.01% 27.34% 8.63%
CU Bird HD-CapsNet $~\dagger$ 106.01 36.59% 17.78% 10.87% 20.29% 26.56% 22.62% 24.09% 6.28%
CU Bird HD-CapsNet $~\ddagger$ 47.56 35.66% 16.98% 2.14% 14.97% 20.86% 17.13% 21.44% 1.55%
Stanford Cars HD-CapsNet 81.17 53.34% 19.52% 14.05% 26.73% 34.69% 29.73% 29.15% 8.13%
Stanford Cars HD-CapsNet $~\dagger$ 81.17 47.50% 16.39% 11.74% 23.56% 31.40% 26.50% 25.76% 6.19%
Stanford Cars HD-CapsNet $~\ddagger$ 25.85 46.01% 12.29% 1.57% 17.10% 24.04% 19.79% 13.60% 0.87%

HERE, $\dagger$ denotes the HD-CapsNet models without the proposed consistency loss (Lc), and $\ddagger$ denotes those without the skip connections between the secondary capsule layers, respectively.

Citation

If you find this repository useful for your research or if it helps in your project, please consider citing it.

Noor, K.T., Robles-Kelly, A., Zhang, L.Y., Bouadjenek, M.R., Luo, W., 2024. A consistency-aware deep capsule network for hierarchical multi-label image classification. Neurocomputing 604, 128376. https://doi.org/10.1016/j.neucom.2024.128376

Bibtex Citation Style:

 @article{noor2024consistency, 
 title={A consistency-aware deep capsule network for hierarchical multi-label image classification},
 author={Noor, Khondaker Tasrif and Robles-Kelly, Antonio and Zhang, Leo Yu and Bouadjenek, Mohamed Reda and Luo, Wei}, 
 journal={Neurocomputing}, 
 volume={604}, 
 ISSN={0925-2312}, 
 DOI={10.1016/j.neucom.2024.128376}, 
 year={2024}, 
 publisher={Elsevier},
 month=nov, 
 pages={128376}, 
 language={en-GB}
 }