Institution profile

Sejong University

Academic institutionasia · kr
Official website
Research library85linked papers
Opportunities0open roles
Selected work

Representative Papers

Deep Perceptual Enhancement for Medical Image Analysis

Apr 19, 2022IEEE journal of biomedical and health informatics

To address perceptual degradations—including low contrast, brightness imbalance, and noise—caused by hardware limitations in medical imaging devices, this work proposes an end-to-end trainable fully convolutional deep network, the first to systematically unify multi-dimensional perceptual enhancement tasks in medical imaging. The method introduces a residual gating mechanism to suppress visual artifacts during enhancement and employs a multi-objective perceptual loss function jointly optimizing PSNR, LPIPS, and DeltaE. Evaluated across multiple medical imaging modalities, the approach achieves PSNR gains of 5.00–7.00 dB and DeltaE reductions of 4.00–6.00, while significantly improving downstream lesion segmentation and classification performance. These results demonstrate both clinical applicability and strong generalization capability.

19 citationsRead paper

DarkDeblur: Learning single-shot image deblurring in low-light condition

Jul 01, 2023Expert systems with applications

This paper addresses the coupled degradation problem of motion blur removal from single-frame images under low-light conditions. We propose the first end-to-end blind deblurring method jointly modeling low illumination and motion blur. Our key contributions are: (1) a lighting-adaptive feature disentanglement module that explicitly separates illumination-variation features from motion-degradation features; (2) a U-Net architecture integrating a physics-based camera response model, differentiable exposure estimation, and multi-scale residual attention mechanisms; and (3) a noise-robust composite loss function. Evaluated on both synthetic and real-world low-light blurred datasets, our method achieves over 2.1 dB PSNR improvement over state-of-the-art deblurring and low-light enhancement methods. Comprehensive qualitative and quantitative results demonstrate superior effectiveness and generalization capability.

14 citations1 influentialRead paper

Two-Stage Deep Denoising With Self-Guided Noise Attention for Multimodal Medical Images

May 01, 2024IEEE Transactions on Radiation and Plasma Medical Sciences

Existing medical image denoising methods suffer from artifact generation and loss of anatomical details when applied across heterogeneous modalities (CT/MRI/US) and diverse noise distributions. To address this, we propose a two-stage deep denoising framework: the first stage estimates residual noise via deep residual learning; the second stage introduces a novel self-guided noise-attention mechanism that explicitly models fine-grained correlations between noise characteristics and input features, integrated with cross-modal feature alignment and joint training to achieve unified adaptation across modalities and noise types. Built upon a cascaded architecture, our method consistently outperforms state-of-the-art approaches across multiple quantitative metrics—achieving improvements of +7.64 in PSNR, +0.1021 in SSIM, −0.80 in DeltaE, +0.1855 in VIFP, and −18.54 in MSE—while effectively suppressing artifacts and preserving pathological structures with high fidelity.

4 citationsRead paper

Illuminating Darkness: Enhancing Real-world Low-light Scenes with Smartphone Images

Mar 10, 2025

To address the performance bottleneck of single-exposure low-light image enhancement caused by the scarcity of high-quality paired data in real-world scenarios, this paper introduces the first large-scale, high-resolution real-world low-light enhancement dataset—comprising 6,425 pairs of 4K+ images captured under illumination levels ranging from 0.1 to 200 lux—and proposes the LC-Tuning Fork Transformer, a luminance-chrominance (LC) decoupled architecture. The model pioneers an LC-decoupled modeling paradigm, incorporating LC cross-attention, LC refinement blocks, and LC-guided supervision to jointly optimize noise suppression and perceptual consistency. Training employs multi-scale non-overlapping patch strategies, while data acquisition and annotation are conducted using dynamic real smartphone sensors. Evaluated on a 400-pair benchmark, our method significantly outperforms state-of-the-art approaches. It demonstrates strong generalization across diverse smartphone hardware and complex indoor/outdoor scenes. Both code and the full dataset are publicly released to advance practical deployment of low-light enhancement.

1 citations1 influentialRead paper

Instruction-Driven 3D Facial Expression Generation and Transition

Jan 13, 2026IEEE transactions on multimedia

This work addresses the limitation of existing methods that typically support only six basic 3D facial expressions, hindering fine-grained and semantically driven generation and transitions. To overcome this, the authors propose the I2FET framework, which enables text-driven synthesis of arbitrary 3D facial expressions and smooth transitions between them. The key innovations include an IFED module for multimodal alignment between textual instructions and facial expression features, and a vertex reconstruction loss to enhance semantic consistency in the latent space. Evaluated on the CK+ and CelebV-HQ datasets, the proposed method significantly outperforms current approaches, generating high-fidelity, semantically accurate, and naturally continuous 3D facial expression sequences.

1 citationsRead paper
Recent publications

Latest Papers

Do People Follow AI Advice? Evidence from a Pension Portfolio Choice Experiment

Aug 11, 2026

This study investigates how individuals respond to AI-generated financial advice in real-world pension investment decisions and its causal impact on asset allocation. Through a 2×2 randomized controlled trial involving 400 Korean workplace pension participants, the research delivered either aggressive or conservative AI recommendations—with or without explanatory rationales—and combined behavioral economics tasks with econometric analysis to identify the causal effects of AI advice in an authentic pension context. The findings reveal that approximately 37% of the recommended portfolio shifts were transmitted to participants’ final allocations, significantly altering expected returns, volatility, and risk profiles. While 81% of participants adjusted their choices—95% of whom moved in the direction of the advice—they implemented only about half of the suggested change on average. Notably, providing explanatory rationales did not significantly enhance compliance. The results highlight selective adoption and partial adherence to AI advice, offering empirical insights for the design of robo-advisory systems.

0 citationsRead paper

Comparative Validation of GPT-4o-mini and Teacher Mean Scores for Automated Scoring of Music Analysis Responses: Single-Pass Deployment, Repeatability, and Strategy-Specific Bias

Aug 03, 2026

This study addresses the challenges of time-intensive and expertise-dependent human scoring in open-ended music analysis assessments by systematically evaluating the feasibility of automated scoring using a large language model (GPT-4o-mini). Leveraging 300 undergraduate responses and teacher-assigned benchmark scores, the authors compare three prompting strategies—few-shot with chain-of-thought (Fs+CoT), retrieval-augmented generation (RAG), and self-consistency (SC)—across four scoring dimensions through multiple experimental rounds to assess stability. Results indicate that Fs+CoT achieves the highest agreement with human raters both in single-run and median-aggregated settings; RAG exhibits a systematic tendency to overrate; SC demonstrates strong repeatability but weaker individual-level consistency; and scoring agreement is consistently lower on terminology than on reasoning dimensions. The findings underscore the critical role of prompting strategies in scoring validity and offer empirical support for automating educational assessment.

0 citationsRead paper