Minhye Chang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Minhye Chang has a total of 31 co-authors. Here are the top co-authors by number of shared publications:
Kye Young Jeong
3 shared publicationsChang Hee Jung
Asan Medical Center, University of Ulsan College of Medicine, Seoul, Republic of Korea
2 shared publications- Wonju Lee
Korea Electrotechnology Research Institute Ansan‐si Gyeonggi‐do South Korea
2 shared publications - Anne Roudaut
Department of Computer Science, University of Bristol, Bristol BS8 1TH, UK
1 shared publication Byung Cheol Song
1 shared publication- Byung Cheol Song
Department of Electrical and Computer Engineering, Inha University, Incheon 22212, Republic of Korea
1 shared publication
Latest Publications
- Open Access
- 2 September 2026
- Published by Elsevier in SSRN Electronic Journal
Towards Efficient Vision-Language Model Inference via Stepwise Visual Token Pruning
Vision-Language Models (VLMs) generate hundreds to thousands of visual tokens per image, making token-dependent computation across decoder layers, including projection, FFN, and attention, as well as KV-cache memory, a major inference bottleneck. Existing token pruning methods address this issue in two ways. Pre-decoder pruning shortens the input sequence but cannot account for the textual query, whereas in-decoder pruning enables query-aware selection but retains the full token cost in early decoder layers. We show that these two strategies are structurally complementary and propose a training-free two-stage framework that exploits this complementarity through stepwise token reduction. In the first stage, CaNS (CLS-Attention Neighbor Selection) performs spatially coherent pre-decoder pruning to shorten the sequence while preserving salient regions. In the second stage, DaFT (Dual-Attention Filtered Token Pruning) refines visual tokens within the decoder usinga V2T→T2V dual-attention mechanism for query-conditioned selection. On LLaVA-1.5, our method reduces 576 visual tokens to 32, a 94.4% reduction, while retaining 97.1% of the baseline performance and reducing FLOPs to 27.9% of the baseline decoder cost, or 37.8% of the full-model cost including the vision encoder. It also generalizes to LLaVA-NeXT, achieving up to a 3.00× speedup under aggressive compression.
- 2 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessTowards Efficient Vision-Language Model Inference via Stepwise Visual Token Pruning
Vision-Language Models (VLMs) generate hundreds to thousands of visual tokens per image, making token-dependent computation across decoder layers, including projection, FFN, and attention, as well as KV-cache memory, a major inference bottleneck. Existing token pruning methods address this issue in two ways. Pre-decoder pruning shortens the input sequence but cannot account for the textual query, whereas in-decoder pruning enables query-aware selection but retains the full token cost in early decoder layers. We show that these two strategies are structurally complementary and propose a training-free two-stage framework that exploits this complementarity through stepwise token reduction. In the first stage, CaNS (CLS-Attention Neighbor Selection) performs spatially coherent pre-decoder pruning to shorten the sequence while preserving salient regions. In the second stage, DaFT (Dual-Attention Filtered Token Pruning) refines visual tokens within the decoder usinga V2T→T2V dual-attention mechanism for query-conditioned selection. On LLaVA-1.5, our method reduces 576 visual tokens to 32, a 94.4% reduction, while retaining 97.1% of the baseline performance and reducing FLOPs to 27.9% of the baseline decoder cost, or 37.8% of the full-model cost including the vision encoder. It also generalizes to LLaVA-NeXT, achieving up to a 3.00× speedup under aggressive compression.
- Open Access
- 1 January 2026
- Published by Institution of Engineering and Technology (IET) in Electronics Letters
- Vol. 62 (1)
Enhancing Endoscopic Artefact Classification With GAN‐Based Synthetic Images Using Artefact Localisation Loss
Accurate artefact classification is essential for reliable diagnosis and computer‐aided analysis in endoscopy, as visual artefacts can severely degrade image quality. However, endoscopic artefact classification is often hindered by limited and imbalanced datasets. This letter introduces an artefact localisation loss (ALL) that guides StyleGAN2 with adaptive discriminator augmentation (StyleGAN2‐ADA) to focus on artefact regions, thereby improving the realism of synthetic images generated from small datasets. The synthetic images were then used for data augmentation, which significantly enhanced classification performance, particularly for underrepresented artefact classes. Experiments on in‐house and endoscopy artefact detection challenge (EAD2020) datasets achieved lower FID scores and showed consistent improvements in F1‐scores, demonstrating the effectiveness of the proposed ALL and synthetic augmentation for robust artefact classification.
- 1 January 2026
- Published by Institution of Engineering and Technology (IET) in Electronics Letters
- Vol. 62 (1)
Open AccessEnhancing Endoscopic Artefact Classification With GAN‐Based Synthetic Images Using Artefact Localisation Loss
Accurate artefact classification is essential for reliable diagnosis and computer‐aided analysis in endoscopy, as visual artefacts can severely degrade image quality. However, endoscopic artefact classification is often hindered by limited and imbalanced datasets. This letter introduces an artefact localisation loss (ALL) that guides StyleGAN2 with adaptive discriminator augmentation (StyleGAN2‐ADA) to focus on artefact regions, thereby improving the realism of synthetic images generated from small datasets. The synthetic images were then used for data augmentation, which significantly enhanced classification performance, particularly for underrepresented artefact classes. Experiments on in‐house and endoscopy artefact detection challenge (EAD2020) datasets achieved lower FID scores and showed consistent improvements in F1‐scores, demonstrating the effectiveness of the proposed ALL and synthetic augmentation for robust artefact classification.
- Open Access
- 22 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 4917-4928
Analysis of the Correlation Between Structure and Hydration Status in Diabetes-Associated and Aged Skin Using Hyperspectral Polarization and SWIR Imaging
We employed hyperspectral polarization imaging and short-wave infrared (SWIR) spectral imaging to characterize dorsal hand skin in young healthy adults and older individuals with type 2 diabetes mellitus (T2DM). Polarization contrast maps (PCM) quantified tissue microstructure, while hydration contrast maps (HCM) assessed water content. Both metrics were analyzed in terms of mean values and spatial complexity using the coefficient of variation and Shannon entropy. Compared to healthy controls, diabetic skin exhibited significantly elevated mean PCM values and increased spatial heterogeneity in both PCM and HCM domains. These findings are consistent with spatially irregular tissue remodeling rather than uniform structural changes. Correlation analysis revealed a negative relationship between PCM and hydration in healthy skin that was absent in diabetic tissue, suggesting a decoupling of structural and hydration regulation in metabolically compromised skin. This dual-domain framework, which integrates mean and spatially resolved optical metrics spanning both structural and functional dimensions, may provide non-invasive optical signatures indicative of skin remodeling associated with aging and diabetes. Such signatures could enable objective disease monitoring and support the evaluation of therapeutic responses.
- 22 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 4917-4928
Open AccessAnalysis of the Correlation Between Structure and Hydration Status in Diabetes-Associated and Aged Skin Using Hyperspectral Polarization and SWIR Imaging
We employed hyperspectral polarization imaging and short-wave infrared (SWIR) spectral imaging to characterize dorsal hand skin in young healthy adults and older individuals with type 2 diabetes mellitus (T2DM). Polarization contrast maps (PCM) quantified tissue microstructure, while hydration contrast maps (HCM) assessed water content. Both metrics were analyzed in terms of mean values and spatial complexity using the coefficient of variation and Shannon entropy. Compared to healthy controls, diabetic skin exhibited significantly elevated mean PCM values and increased spatial heterogeneity in both PCM and HCM domains. These findings are consistent with spatially irregular tissue remodeling rather than uniform structural changes. Correlation analysis revealed a negative relationship between PCM and hydration in healthy skin that was absent in diabetic tissue, suggesting a decoupling of structural and hydration regulation in metabolically compromised skin. This dual-domain framework, which integrates mean and spatially resolved optical metrics spanning both structural and functional dimensions, may provide non-invasive optical signatures indicative of skin remodeling associated with aging and diabetes. Such signatures could enable objective disease monitoring and support the evaluation of therapeutic responses.
- Open Access
- 19 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 51911-51921
Real-Time Long-Wave Infrared Semantic Segmentation With Adaptive Noise Reduction and Feature Fusion
Semantic segmentation in the Long-Wave Infrared (LWIR) domain is critical for a wide range of applications including emergency response, industrial safety monitoring, and public building security. However, LWIR images often suffer from inherent pattern noise and ambiguous object boundaries that hinder accurate segmentation. To address these challenges, this study presents a novel real-time semantic segmentation framework specifically designed for LWIR images. The framework incorporates a Stripe noise Denoising Module (SDM) and a Boundary Enhancement Module (BEM), which leverage frequency-domain filtering and learnable weights to adaptively process noisy input data and improve boundary prediction. In addition, a Multi-Stream Fusion Module (MSFM) integrates multi-scale semantic features with boundary information, thereby enhancing segmentation accuracy across diverse object scales. The proposed method demonstrates state-of-the-art performance in both accuracy and efficiency on multiple datasets, including KERInha and SODA. Extensive qualitative and quantitative evaluations further validate its robustness, particularly in scenarios where RGB imagery is unavailable. By eliminating the need for supplementary information such as depth data, this approach facilitates precise indoor and outdoor segmentation tasks with lightweight computation, making it highly suitable for real-world applications.
- 19 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 51911-51921
Open AccessReal-Time Long-Wave Infrared Semantic Segmentation With Adaptive Noise Reduction and Feature Fusion
Semantic segmentation in the Long-Wave Infrared (LWIR) domain is critical for a wide range of applications including emergency response, industrial safety monitoring, and public building security. However, LWIR images often suffer from inherent pattern noise and ambiguous object boundaries that hinder accurate segmentation. To address these challenges, this study presents a novel real-time semantic segmentation framework specifically designed for LWIR images. The framework incorporates a Stripe noise Denoising Module (SDM) and a Boundary Enhancement Module (BEM), which leverage frequency-domain filtering and learnable weights to adaptively process noisy input data and improve boundary prediction. In addition, a Multi-Stream Fusion Module (MSFM) integrates multi-scale semantic features with boundary information, thereby enhancing segmentation accuracy across diverse object scales. The proposed method demonstrates state-of-the-art performance in both accuracy and efficiency on multiple datasets, including KERInha and SODA. Extensive qualitative and quantitative evaluations further validate its robustness, particularly in scenarios where RGB imagery is unavailable. By eliminating the need for supplementary information such as depth data, this approach facilitates precise indoor and outdoor segmentation tasks with lightweight computation, making it highly suitable for real-world applications.
- Open Access
- 16 October 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 153822-153831
Efficient Mapping of Tissue Oxygen Saturation Using Hyperspectral Imaging and GAN
Oxygen saturation is a crucial metric used in various medical contexts, indicating the concentration of oxygen in the blood. Recently, hyperspectral imaging-based oximetry techniques have been proposed for broad applications of two-dimensional information without sensor contact. However, these methods suffer from high computational complexity and inefficiency due to large hypercubes. In this study, we investigate an efficient tissue oxygen saturation (StO2) mapping from hyperspectral image using a generative adversarial network (GAN). The model was trained on hyperspectral images acquired from 40 healthy individuals. The Non-dominated Sorting Genetic Algorithm II (NSGA-II) was then used to identify the optimal combination of hyperspectral bands to reduce the number of bands required for computation while maintaining performance. The selected optimal band combinations included both visible and near-infrared (NIR) bands. The GAN-based model with the ten-band combination estimated StO2 effectively, achieving a root mean squared error (RMSE) of 0.1169 ± 0.0350. Finally, the model trained on healthy subjects’ data was applied to the palm and sole images of 20 patients with type 2 diabetes mellitus (DM) to validate its performance. The performance on the patient data was comparable to that obtained from healthy individuals, with the best performance achieved using the ten-band combination (RMSE=0.1077 ± 0.0341). In conclusion, the developed efficient hyperspectral image-based oxygenation mapping method demonstrated robust performance for patients with type 2 DM as well as healthy subjects.
- 16 October 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 153822-153831
Open AccessEfficient Mapping of Tissue Oxygen Saturation Using Hyperspectral Imaging and GAN
Oxygen saturation is a crucial metric used in various medical contexts, indicating the concentration of oxygen in the blood. Recently, hyperspectral imaging-based oximetry techniques have been proposed for broad applications of two-dimensional information without sensor contact. However, these methods suffer from high computational complexity and inefficiency due to large hypercubes. In this study, we investigate an efficient tissue oxygen saturation (StO2) mapping from hyperspectral image using a generative adversarial network (GAN). The model was trained on hyperspectral images acquired from 40 healthy individuals. The Non-dominated Sorting Genetic Algorithm II (NSGA-II) was then used to identify the optimal combination of hyperspectral bands to reduce the number of bands required for computation while maintaining performance. The selected optimal band combinations included both visible and near-infrared (NIR) bands. The GAN-based model with the ten-band combination estimated StO2 effectively, achieving a root mean squared error (RMSE) of 0.1169 ± 0.0350. Finally, the model trained on healthy subjects’ data was applied to the palm and sole images of 20 patients with type 2 diabetes mellitus (DM) to validate its performance. The performance on the patient data was comparable to that obtained from healthy individuals, with the best performance achieved using the ten-band combination (RMSE=0.1077 ± 0.0341). In conclusion, the developed efficient hyperspectral image-based oxygenation mapping method demonstrated robust performance for patients with type 2 DM as well as healthy subjects.
- Open Access
Outdoor Localization Using BLE RSSI and Accessible Pedestrian Signals for the Visually Impaired at Intersections
One of the major challenges for blind and visually impaired (BVI) people is traveling safely to cross intersections on foot. Many countries are now generating audible signals at crossings for visually impaired people to help with this problem. However, these accessible pedestrian signals can result in confusion for visually impaired people as they do not know which signal must be interpreted for traveling multiple crosses in complex road architecture. To solve this problem, we propose an assistive system called CAS (Crossing Assistance System) which extends the principle of the BLE (Bluetooth Low Energy) RSSI (Received Signal Strength Indicator) signal for outdoor and indoor location tracking and overcomes the intrinsic limitation of outdoor noise to enable us to locate the user effectively. We installed the system on a real-world intersection and collected a set of data for demonstrating the feasibility of outdoor RSSI tracking in a series of two studies. In the first study, our goal was to show the feasibility of using outdoor RSSI on the localization of four zones. We used a k-nearest neighbors (kNN) method and showed it led to 99.8% accuracy. In the second study, we extended our work to a more complex setup with nine zones, evaluated both the kNN and an additional method, a Support Vector Machine (SVM) with various RSSI features for classification. We found that the SVM performed best using the RSSI average, standard deviation, median, interquartile range (IQR) of the RSSI over a 5 s window. The best method can localize people with 97.7% accuracy. We conclude this paper by discussing how our system can impact navigation for BVI users in outdoor and indoor setups and what are the implications of these findings on the design of both wearable and traffic assistive technology for blind pedestrian navigation.
Open AccessOutdoor Localization Using BLE RSSI and Accessible Pedestrian Signals for the Visually Impaired at Intersections
One of the major challenges for blind and visually impaired (BVI) people is traveling safely to cross intersections on foot. Many countries are now generating audible signals at crossings for visually impaired people to help with this problem. However, these accessible pedestrian signals can result in confusion for visually impaired people as they do not know which signal must be interpreted for traveling multiple crosses in complex road architecture. To solve this problem, we propose an assistive system called CAS (Crossing Assistance System) which extends the principle of the BLE (Bluetooth Low Energy) RSSI (Received Signal Strength Indicator) signal for outdoor and indoor location tracking and overcomes the intrinsic limitation of outdoor noise to enable us to locate the user effectively. We installed the system on a real-world intersection and collected a set of data for demonstrating the feasibility of outdoor RSSI tracking in a series of two studies. In the first study, our goal was to show the feasibility of using outdoor RSSI on the localization of four zones. We used a k-nearest neighbors (kNN) method and showed it led to 99.8% accuracy. In the second study, we extended our work to a more complex setup with nine zones, evaluated both the kNN and an additional method, a Support Vector Machine (SVM) with various RSSI features for classification. We found that the SVM performed best using the RSSI average, standard deviation, median, interquartile range (IQR) of the RSSI over a 5 s window. The best method can localize people with 97.7% accuracy. We conclude this paper by discussing how our system can impact navigation for BVI users in outdoor and indoor setups and what are the implications of these findings on the design of both wearable and traffic assistive technology for blind pedestrian navigation.
- Open Access
- 16 June 2019
- Published by Wiley in Computational Intelligence and Neuroscience
- Vol. 2019, 1-12
Enhancing the Usability of Brain-Computer Interface Systems
Brain-computer interfaces (BCIs) aim to enable people to interact with the external world through an alternative, nonmuscular communication channel that uses brain signal responses to complete specific cognitive tasks. BCIs have been growing rapidly during the past few years, with most of the BCI research focusing on system performance, such as improving accuracy or information transfer rate. Despite these advances, BCI research and development is still in its infancy and requires further consideration to significantly affect human experience in most real-world environments. This paper reviews the most recent studies and findings about ergonomic issues in BCIs. We review dry electrodes that can be used to detect brain signals with high enough quality to apply in BCIs and discuss their advantages, disadvantages, and performance. Also, an overview is provided of the wide range of recent efforts to create new interface designs that do not induce fatigue or discomfort during everyday, long-term use. The basic principles of each technique are described, along with examples of current applications in BCI research. Finally, we demonstrate a user-friendly interface paradigm that uses dry capacitive electrodes that do not require any preparation procedure for EEG signal acquisition. We explore the capacitively measured steady-state visual evoked potential (SSVEP) response to an amplitude-modulated visual stimulus and the auditory steady-state response (ASSR) to an auditory stimulus modulated by familiar natural sounds to verify their availability for BCI. We report the first results of an online demonstration that adopted this ergonomic approach to evaluating BCI applications. We expect BCI to become a routine clinical, assistive, and commercial tool through advanced EEG monitoring techniques and innovative interface designs.
- 16 June 2019
- Published by Wiley in Computational Intelligence and Neuroscience
- Vol. 2019, 1-12
Open AccessEnhancing the Usability of Brain-Computer Interface Systems
Brain-computer interfaces (BCIs) aim to enable people to interact with the external world through an alternative, nonmuscular communication channel that uses brain signal responses to complete specific cognitive tasks. BCIs have been growing rapidly during the past few years, with most of the BCI research focusing on system performance, such as improving accuracy or information transfer rate. Despite these advances, BCI research and development is still in its infancy and requires further consideration to significantly affect human experience in most real-world environments. This paper reviews the most recent studies and findings about ergonomic issues in BCIs. We review dry electrodes that can be used to detect brain signals with high enough quality to apply in BCIs and discuss their advantages, disadvantages, and performance. Also, an overview is provided of the wide range of recent efforts to create new interface designs that do not induce fatigue or discomfort during everyday, long-term use. The basic principles of each technique are described, along with examples of current applications in BCI research. Finally, we demonstrate a user-friendly interface paradigm that uses dry capacitive electrodes that do not require any preparation procedure for EEG signal acquisition. We explore the capacitively measured steady-state visual evoked potential (SSVEP) response to an amplitude-modulated visual stimulus and the auditory steady-state response (ASSR) to an auditory stimulus modulated by familiar natural sounds to verify their availability for BCI. We report the first results of an online demonstration that adopted this ergonomic approach to evaluating BCI applications. We expect BCI to become a routine clinical, assistive, and commercial tool through advanced EEG monitoring techniques and innovative interface designs.
- 10 December 1991
- Published by Wiley in ChemInform
- Vol. 22 (49)
ChemInform Abstract: Intramolecular Reductive Cyclization of Aldehydes and Ketones with Alkynes Promoted by Samarium(II) Iodide.
ChemInform is a weekly Abstracting Service, delivering concise information at a glance that was extracted from about 100 leading journals. To access a ChemInform Abstract of an article which was published elsewhere, please select a “Full Text” option. The original article is trackable via the “References” option.
- 10 December 1991
- Published by Wiley in ChemInform
- Vol. 22 (49)
ChemInform Abstract: Intramolecular Reductive Cyclization of Aldehydes and Ketones with Alkynes Promoted by Samarium(II) Iodide.
ChemInform is a weekly Abstracting Service, delivering concise information at a glance that was extracted from about 100 leading journals. To access a ChemInform Abstract of an article which was published elsewhere, please select a “Full Text” option. The original article is trackable via the “References” option.