Journal article · 2025 · Open access

Enhancing Vision Transformer Performance Using Superpixel Tokenization

Jihun Yang, Dong Joong Kang

Journal of Korea Multimedia Society, 28(11), 1831–1840 · Published 30 November 2025 · DOI 10.9717/kmms.2025.28.11.1831

Open full textCite this paper

In brief

“Enhancing Vision Transformer Performance Using Superpixel Tokenization” is a 2025 journal article by Jihun Yang and Dong Joong Kang, published in Journal of Korea Multimedia Society. OpenAlex counts 0 citations as of 22 November 2026. It is open access: the full text is free to read at Journal of Korea Multimedia Society.

Citations
0
Authors
2

Abstract

Vision transformers have demonstrated remarkable performance across various computer vision tasks, owing to their self-attention mechanism, which effectively models long-range dependencies among tokens. However, the presence of mixed local representations within each token poses a challenge, limiting the ability of self-attention to learn a well-structured global representation. In this work, we propose a superpixel-based tokenization approach for vision transformers, which standardizes the semantic characteristics of each token to preserve local information while facilitating more effective global representation learning. This method, referred to as SuperPixel Tokenization (SPT), consists of two key components: first, patch packaging, which crops each superpixel to its minimum bounding rectangle, zero-pads non-superpixel regions, and resizes it to a fixed-size square patch to standardize token representation; second, a token fusion strategy that integrates token embeddings from the original and superpixel-based images using a gated mechanism. Our approach improves Top-1 classification accuracy by 1.63% on ImageNet-1K, 5.39% on CIFAR-10, and 8.50% on CIFAR-100 compared to the baseline model. Furthermore, attention map analysis reveals that our method enhances global representation learning in the shallow layers.

Cite this paper

Formatted from the catalogue record. Check names, capitals and page numbers against the paper before you submit.

APA 7

Yang, J., & Kang, D. J. (2025). Enhancing Vision Transformer Performance Using Superpixel Tokenization. Journal of Korea Multimedia Society, 28(11), 1831–1840. https://doi.org/10.9717/kmms.2025.28.11.1831

MLA 9

Yang, Jihun, and Dong Joong Kang. “Enhancing Vision Transformer Performance Using Superpixel Tokenization.” Journal of Korea Multimedia Society, vol. 28, no. 11, 2025, pp. 1831–1840. https://doi.org/10.9717/kmms.2025.28.11.1831.

Chicago

Yang, Jihun and Dong Joong Kang. 2025. “Enhancing Vision Transformer Performance Using Superpixel Tokenization.” Journal of Korea Multimedia Society 28 (11): 1831–1840. https://doi.org/10.9717/kmms.2025.28.11.1831.

Harvard

Yang, J. and Kang, D. J. (2025) ‘Enhancing Vision Transformer Performance Using Superpixel Tokenization’, Journal of Korea Multimedia Society, 28(11), pp. 1831–1840. doi:10.9717/kmms.2025.28.11.1831.

IEEE

J. Yang and D. J. Kang, “Enhancing Vision Transformer Performance Using Superpixel Tokenization,” Journal of Korea Multimedia Society, vol. 28, no. 11, pp. 1831–1840, 2025, doi: 10.9717/kmms.2025.28.11.1831.

BibTeX

@article{yang2025enhancing,
  title     = {{Enhancing Vision Transformer Performance Using Superpixel Tokenization}},
  author    = {Yang, Jihun and Kang, Dong Joong},
  journal   = {Journal of Korea Multimedia Society},
  year      = {2025},
  volume    = {28},
  number    = {11},
  pages     = {1831--1840},
  doi       = {10.9717/kmms.2025.28.11.1831},
  url       = {https://doi.org/10.9717/kmms.2025.28.11.1831}
}

Download for a reference manager: BibTeX (.bib) · RIS for Zotero, Mendeley and EndNote (.ris)

Topic and keywords

Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications

  • vision transformer
  • self-attention
  • token embeddings
  • gating mechanism
  • image classification
  • attention maps

Work with this paper

Most-cited papers in Computer Science

All highly cited Computer Science papers

Questions about this paper

How do I cite this paper in APA style?

Yang, J., & Kang, D. J. (2025). Enhancing Vision Transformer Performance Using Superpixel Tokenization. Journal of Korea Multimedia Society, 28(11), 1831–1840. https://doi.org/10.9717/kmms.2025.28.11.1831

How many times has this paper been cited?

0 times, according to OpenAlex on 22 November 2026. Google Scholar and Scopus count different sets of sources, so their numbers differ.

Is this paper free to read?

Yes. It is open access at Journal of Korea Multimedia Society: sign in and open the full text from this page.

Who wrote this paper?

Jihun Yang, Dong Joong Kang.

What is the DOI of this paper?

10.9717/kmms.2025.28.11.1831, permanently at https://doi.org/10.9717/kmms.2025.28.11.1831. Sign in and open it from this page.

Bibliographic data from OpenAlex (CC0), last checked 22 November 2026. ResearchGuru does not host this paper; once you are signed in, the links open it at the publisher or the repository that does.