Ph.D. Student · Korea University

Sejin Park

Department of Electrical Engineering
Korea University, Seoul, Korea

Efficient AI Low-level Vision

I am a Ph.D. student in Electrical Engineering at Korea University, advised by Prof. Seung-Won Jung. My research focuses on efficient AI and low-level vision, with an emphasis on inference-efficient and hardware-aware models, including logic-based neural networks, image restoration, and camera ISP.

Email Google Scholar GitHub LinkedIn CV

Research Interests

Efficient AI

Inference-efficient models, data-efficient learning, hardware-aware models, and logic-based neural networks.

Low-level Vision

Image restoration and camera ISP, with an emphasis on computationally efficient image processing.

Selected Papers (First-Author)

2026 BK21 FOUR Top-Tier Conference

Overcoming Kernel Redundancy for
Scaling Logic Gate Networks

Can logic gate networks turn width into effective capacity?

Observation

A function-aware tree edit distance reveals increasing kernel redundancy with width, especially at the first gate level.

Idea

We propose DLK for input-dependent routing across logic kernel groups, and EDLK to focus routing on the highly redundant first gate level.

Outcome

Turns additional width into effective capacity, improving accuracy and parameter efficiency.

Sejin Park, Hongjae Lee, Changwoo Han, Seung-Won Jung

Conference on Neural Information Processing Systems (NeurIPS 2026)

Early-stage Dynamic Logic Kernel overview

2025 BK21 FOUR Top-Tier Conference

IM-LUT: Interpolation Mixing Look-Up Tables for Image Super-Resolution

Can LUTs alone enable arbitrary-scale super-resolution?

Idea

Learn interpolation mixing weights and convert them into LUTs for efficient arbitrary-scale super-resolution.

Outcome

Enables continuous-scale super-resolution while retaining efficient LUT-based inference.

Sejin Park, Sangmin Lee, Kyong Hwan Jin, Seung-Won Jung

IEEE/CVF International Conference on Computer Vision (ICCV 2025)

Paper Code Video

2023 BK21 FOUR Top-Tier Conference

FDCNet: Feature Drift Compensation Network for Class-Incremental Weakly Supervised Object Localization

Can weakly supervised localization learn new classes without forgetting previously learned localization knowledge?

Idea

Compensate feature drift caused by class-incremental learning to preserve localization capability across sequential tasks.

Outcome

Maintains localization knowledge while progressively learning new object classes.

Sejin Park*, Taehyung Lee*, Yeejin Lee, Byeongkeun Kang

ACM International Conference on Multimedia (ACM MM 2023)

* Equal contribution

Paper Code

Under Review Two manuscripts are currently under review.

Manuscript 1 NRF Student Research Grant

Juhyun Bae*, Sejin Park*, Seung-Won Jung

* Equal contribution

Manuscript 2

Sejin Park, Soeun Lee, Taehyung Lee, Yeejin Lee, Byeongkeun Kang

Book

2025

Introduction to Computer Vision for Future Mobility

Sejin Park and Seung-Won Jung

ISBN: 9791175000766

Patent Application

2025

Lookup table-based multi-interpolation hybrid method and device for arbitrary scale image super-resolution

Sejin Park and Seung-Won Jung

KR Application 10-2025-0104234