Research

Themes

Low-light image enhancement

Based on Retinex theory, I study methods that brighten images taken in dark places. After Fast Single-Scale Retinex and XCR, which uses a complex exponential kernel, I published SD-Retinex, designed around the derivative of the sigmoid function.

In an evaluation on the LOL dataset (eval15), SD-Retinex gave the best results among training-free methods that run on a CPU.

A higher PSNR is better, and a smaller color difference ΔE is better. Processing time is per image on a CPU (AMD Ryzen 7 8840U). Without any training, it also surpassed the zero-reference deep learning methods Zero-DCE and SCI in PSNR and ΔE. The code is available on GitHub.

Kyudo shooting form evaluation with real-time pose estimation

We proposed a method that estimates the skeleton in real time to evaluate kyudo (Japanese archery) form, and implemented it on the web.

Regional policy analysis with small language models

We analyzed 1,721 municipalities with 125 indicators (“UBRS,” built on UMAP, HDBSCAN and SHAP). We proposed a framework that gives the features of the resulting clusters to the small language model Qwen2.5-14B-Instruct (4-bit) through RAG to obtain policy insights. It runs on consumer hardware, and overall relevance improved by 16.0 points compared with a persona-only baseline (Wilcoxon test, p < 1e-10).

Entrepreneurship education

I propose an entrepreneurship education model for KOSEN students that is based on failure cases.

Papers and presentations

International conference (peer-reviewed)

• Toma Okugawa, Keito Inoshita. “Re-Defining Vanishing Municipalities in Japan: A Multidimensional Clustering and SLM-Based Policy Insight Framework.” Proceedings of the IEEE 15th Global Conference on Consumer Electronics (IEEE GCCE 2026), to appear. Oral presentation, Kobe International Conference Center, October 2026.

Journal and domestic conference papers

• Takahiro Makiyama, Haruki Ishibashi, Toma Okugawa, Yasuyuki Kubo. “Educational practice at university using Huygens’ cycloidal pendulum clock” (in Japanese). Journal of the Physics Education Society of Japan (物理教育), vol. 74, no. 3, September 2026.
• Atsuki Seo, Hiroyoshi Enomoto, Haruki Ishibashi, Toma Okugawa, Shosei Hamada. “KiWi: A Secure Information Request and Management System for Public Institutions through the Digitization of Paper-Based Personal Data” (in Japanese). Proceedings of the IPSJ National Convention, February 20, 2026.
• Toma Okugawa, Tomonari Masuzaki, Takahiro Makiyama. “Proposal and Web Implementation of a Kyudo Form Evaluation Method Based on Real-Time Pose Estimation” (in Japanese). Shikoku-section Joint Convention Record of the Institutes of Electrical and Related Engineers, September 22, 2025. Oral presentation on September 27, 2025; Outstanding Presentation Award.

Preprints

• Toma Okugawa. “Recognition Without Revenue: Designing a Failure-Case-Based Entrepreneurship Education Model for Japanese KOSEN Students.” EdArXiv, September 12, 2026.
  doi.org/10.35542/osf.io/dr9jx_v1
• Toma Okugawa. “SD-Retinex: A Stable Retinex Method for Low-Light Image Enhancement Designed Around the Sigmoid Derivative.” Zenodo, June 27, 2026.
  doi.org/10.5281/zenodo.20945211
  Code: github.com/toma-okugawa/sd-retinex
• Toma Okugawa. “A High-Speed and Structure-Preserving Retinex Method Based on a Complex Exponential Kernel: XCR” (in Japanese). Jxiv, November 20, 2025.
  doi.org/10.51094/jxiv.1961
• Toma Okugawa. “Proposal of Fast Single-Scale Retinex and Evaluation of Real-Time Human Pose Estimation Performance under Low-Light Conditions” (in Japanese). Jxiv, November 12, 2025.
  doi.org/10.51094/jxiv.1897

A list of my work is also available on researchmap.