GIST Scholar
Not a member yet
30271 research outputs found
Sort by
ROS-Triggered Microgels for Programmable Drug Release in Volumetric Muscle Loss Repair
No full textVolumetric muscle loss (VML), frequently resulting from traumatic or surgical damage, causes significant muscle mass depletion and fibrosis, and presents major challenges to effective regeneration. In this study, reduced graphene-containing hyaluronic acid microgels (rGHMs) are developed as a multifunctional platform for reactive oxygen species (ROS)-scavenging and ROS-responsive drug delivery to treat VML. The rGHMs are synthesized via a water-in-oil emulsion and subsequent chemical reduction. rGHMs demonstrate superior antioxidative capability, curcumin loading efficiency, and ROS-mediated drug release properties compared to HA microgels (HMs) and unreduced graphene-oxide containing HA microgels (GHMs). In particular, curcumin-loaded rGHMs (Cur/rGHMs) significantly facilitate curcumin release with a 2.6-fold increase under 1 mm H2O2 compared to non-ROS conditions, demonstrating programmable, ROS-triggered release kinetics. In vitro studies confirm that rGHMs are cytocompatible and protect C2C12 myoblasts from ROS-induced damage. In vivo studies using a mouse VML model reveal that Cur/rGHMs significantly enhance skeletal muscle regeneration, as evidenced by an increased number of centronucleated muscle cells, 89.0 % muscle strength recovery, 51.0 % reduction in fibrosis, a 2.3-fold increase in vascularization, and attenuated inflammatory macrophage infiltration. These ROS-responsive microgels enable programmed curcumin delivery in oxidative environments, offering a promising therapeutic strategy for skeletal muscle regeneration in VML.FALSEsciescopu
A Deep Learning-Based Approach for Early Alzheimer's Disease Diagnosis: Addressing Generalization and Domain Shift Across Populations
No full textAlzheimer’s Disease (AD) remains a leading cause of cognitive decline and dementia worldwide, affecting millions and placing a heavy burden on healthcare systems and caregivers. As the disease progresses over time, it often goes undiagnosed until moderate or late stages, where therapeutic interventions have diminished effectiveness. Timely diagnosis remains the single most impactful intervention, capable of significantly delaying disease progression. However, current diagnostic approaches largely identify AD only in its later stages, missing the critical opportunity for preventive action during the mild cognitive impairment (MCI) phase—a transitional stage where intervention is most beneficial. Therefore, advances in neuroimaging and artificial intelligence, particularly deep learning, have created potential opportunities for early detection of AD and prediction of MCI progression. This thesis explores deep learning as a promising approach in the early detection and progression prediction of Alzheimer’s Disease using structural magnetic resonance imaging (sMRI), addressing key limitations in interpretability, generalization, and data heterogeneity across medical cohorts. Chapter 1 begins by outlining the clinical and neuropathological basis of Alzheimer’s Disease, underscoring the importance of early detection through biomarkers such as hippocampal atrophy, amyloid-beta plaque accumulation, and tau protein tangles. It also introduces recent progress in neuroimaging-based diagnosis and the background of artificial intelligence, particularly deep learning. Structural Magnetic Resonance Imaging (MRI), with its high spatial resolution and soft tissue contrast, enables visualization of cerebral atrophy with significant potential for early diagnosis. Deep learning models, specifically convolutional neural networks (CNNs), have shown remarkable performance in disease classification tasks. This chapter sets the foundation for the subsequent work by building upon and expanding state-of-the-art techniques. Chapter 2 introduces the application of Vision Transformers (ViT) in modeling MCI-to-AD progression, as explored in the study “Vision Transformers for the Prediction of Mild Cognitive Impairment to Alzheimer’s Disease Progression using Mid-Sagittal sMRI.” Departing from CNN-based architectures, this work embraces ViTs for their ability to model global image dependencies using self-attention mechanisms. The model utilizes mid-sagittal slices from structural MRI, targeting brain regions such as the thalamus, medial frontal cortex, and occipital lobe—areas known to be pathologically relevant in MCI conversion. Using data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), the proposed ViT framework achieves high accuracy (83.27%), specificity (85.07%), and sensitivity (81.48%) in distinguishing progressive MCI (pMCI) from stable MCI (sMCI). Importantly, attention maps derived from the transformer layers illuminate which brain regions most influence prediction, thereby enhancing model interpretability. This chapter illustrates the viability of attention-driven architectures in early AD diagnosis. Chapter 3 addresses a critical challenge in AD-related AI studies: data leakage due to improper splitting of training and test datasets. The study “Deep Learning with Guided Attention for Early Diagnosis of Alzheimer’s Disease” introduces the Guided-Attention Deep Learning (GADL) model that explicitly avoids this issue by enforcing subject-level data separation and integrates domain-specific attention with a CNN backbone. The GADL framework leverages a two-part architecture: a guided-attention feature extractor pretrained on the AD vs CN task, and a task-specific classifier for CN vs MCI and pMCI vs sMCI. The innovation lies in its guided-attention mechanism, which leverages knowledge distilled from a CN vs. AD classifier to identify the most informative regions for MCI classification. With subject-level splitting to prevent data leakage, the model achieved 80.29% accuracy for MCI progression prediction on the ADNI dataset and demonstrated high generalizability with 79.38% accuracy on the AIBL dataset. Chapter 4 expands upon the limitations of generalization posed by domain shift—variations in imaging protocols, demographics, and scanner settings that often degrade model performance when deployed across national cohorts. The third study, Domain Adaptation Network with Guided Attention for Multi- Cohort Alzheimer’s Disease Prediction proposes a semi-supervised domain adaptation technique that leverages guided attention scores from AD vs. CN classification to facilitate adaptation between source and target domains. The proposed architecture combines multi-plane sMRI feature extraction (sagittal, axial, and coronal) with semi-supervised domain adaptation and guided attention. Guided attention ensures that domain adaptation focuses on critical brain regions, while adversarial and pseudo-labeling strategies align feature distributions. The model is evaluated on both the ADNI and a separate national cohort with heterogeneous imaging settings, outperforming state-of-the-art baselines. The attention maps generated reveal consistent pathological focus areas across domains, underscoring the model’s robustness and reliability. Chapter 5 presents the most comprehensive and generalizable framework in this thesis: AttCORAL – Domain-Adaptive Attention Networks for Early Alzheimer’s Disease Diagnosis. Building upon prior efforts, this study incorporates Correlation Alignment (CORAL) loss with dual attention mechanisms in a residual neural network (ResNet) architecture. This approach combines the strengths of attention-based deep learning with statistical domain alignment through Correlation Alignment (CORAL) loss. The model architecture builds upon ResNet backbones enhanced with dual attention—spatial and channel— and integrates CORAL loss at the feature level to align the source and target domain distributions. Evaluated across both ADNI and AIBL datasets, AttCORAL achieves superior performance in both CN vs MCI classification and MCI-to-AD prediction tasks. What sets this model apart is its interpretability and adaptability: attention maps visualize the hippocampus, entorhinal cortex, and temporal lobe as key contributors to diagnosis, while CORAL ensures consistent performance across imaging settings and populations. The success of AttCORAL validates the thesis’ central argument—that attention-guided, domain-adaptive learning can lead to robust, interpretable, and generalizable AI models for neurodegenerative disease diagnosis. In summary, this thesis presents a cohesive, iterative exploration of deep learning methods tailored to the early diagnosis of Alzheimer’s Disease with a specific focus on MCI progression prediction. Each successive chapter builds upon the limitations of the previous, gradually improving accuracy, interpretability, and domain generalization. It demonstrates how attention mechanisms, domain adaptation, and rigorous validation protocols can collectively lead to high-performance, clinically meaningful, and ethically robust diagnostic models. As such, the work contributes significantly to the growing field of AI-driven neurodiagnostics and brings us closer to scalable, trustworthy tools for the early management of Alzheimer’s Disease.DoctorAbstract i
Acknowledgments v
List of Figure ix
List of Table xi
I. Background 1
I.1. Alzheimer’s Disease Background 1
I.2. Deep Learning Background 4
II. Vision Transformers for MCI to AD Prediction 9
II.1. Introduction 9
II.2. Materials and methods 12
II.3. Results and analysis 18
II.4. Discussion 23
II.5. Conclusion 25
III. Guided-Attention For Early AD Diagnosis 26
III.1. Introduction 26
III.2. Material and methods. 28
III.3. Results and analysis 36
III.4. Discussion 41
III.5. Conclusion 44
IV. Domain Adaptation with Guided Attention 45
IV.1. Introduction 45
IV.2. Domain Adaptation in Medical Field 46
IV.3. Materials and methods 48
IV.4. Results and analysis 52
IV.5. Conclusion 56
V. AttCORAL Domain-Adaptive Attention Networks 57
V.1. Introduction 57
V.2. Materials and methods 58
V.3. Results and analysis 63
V.4. Discussion 66
V.5. Conclusion 69
VI. Conclusion and Future Work 70
Reference 7
In situ generated diphenylphosphine-chelated imidazo[1,5-a]pyridin-3-ylidene nickel(0) catalysts for highly efficient acrylate synthesis from ethylene and CO2
No full textDiphenylphosphine-chelated imidazo[1,5-a]pyridin-3-ylidene (PPh₂-ImPy) nickel(0) catalysts were synthesized in-situ from imidazolium salts and various commercially available nickel(0) complexes using a base. These catalysts exhibited a highly planar and rigid structure, as confirmed by X-ray crystallographic analysis. Remarkably, the in-situ generated PPh₂-ImPy-Ni(0) catalyst demonstrated a turnover number (TON) of 570 while maintaining a high yield of 82 %, overcoming the typical trade-off between TON and yield often observed in acrylate synthesis reactions, where the use of excess base to achieve high TON typically results in lower product yields. Moreover, the ligand precursors (PPh₂-ImPy∙HCl salts) were stable under open-air conditions, making them easy to handle. The in-situ generation protocol using the ligand precursors and commercially available metal sources, eliminates the need for synthesizing sensitive Ni(0) catalysts. Furthermore, the sodium acrylate product was efficiently isolated from the reaction mixture through a straightforward extraction process. © 2024 The AuthorsTRUEsciescopu
Thermodynamic and Kinetic Investigations of Clathrate Hydrates: Applications to Energy Gas Storage and Separation
No full textClathrate hydrates are crystalline inclusion compounds in which a hydrogen-bonded water lattice encapsulates guest molecules, typically light gases or small organic compounds. Understanding their thermodynamic stability and kinetic behavior is crucial for optimizing their applications in energy gas storage and separation. This dissertation investigates the thermodynamic and kinetic properties of various clathrate hydrate systems, focusing on their potential for methane and hydrogen storage, as well as methane enrichment from gas mixtures. By integrating phase equilibrium measurements, crystallographic analysis, and spectroscopic techniques, this work provides a comprehensive understanding of the molecular interactions governing hydrate formation for Energy gas storage and separation. The dissertation is divided into two main parts. Part 1 explores thermodynamic and kinetic strategies for enhancing hydrate-based natural gas storage. Chapter 2 examines the role of cyclopentyl amine (CPA) as a tunable thermodynamic promoter for methane hydrates, characterizing its impact on phase stability and cage occupancy. Chapter 3 extends this approach to multicomponent natural gas storage, where epoxycyclopentane (ECP) is evaluated for its ability to modulate hydrate formation kinetics and increase gas storage capacity. The findings highlight the importance of molecular tuning for optimizing hydrate-based storage systems. Part 2 investigates hydrate-based gas separation, focusing on methane enrichment from hydrogen-natural gas blends. Chapter 4 assesses the stabilization of hydrogen hydrates using gas-phase thermodynamic promoters, elucidating their effects on hydrate phase behavior and guest molecule distribution. Chapter 5 examines selective methane enrichment via hydrate formation, demonstrating how tailored hydrate structures can enhance methane selectivity over hydrogen. These results provide a foundation for developing efficient hydrate-based gas separation technologies. In conclusion, this dissertation presents strategies based on the unique cage occupation behavior of guest molecules, coupled with the thermodynamic and kinetic properties of clathrate hydrates for gas storage and separation applications. The findings emphasize the significance of leveraging intrinsic hydrate properties to enhance gas storage capacity and selectivity, offering insights for the advancement of hydrate-based technologies in energy and environmental applications.DoctorAbstract i
Contents ii
List of Figures iv
List of Tables vii
Chapter 1. Introduction 1
1.1. General introduction to clathrate hydrates 1
1.1.1. Physical and structural characteristics of clathrate hydrates 1
1.1.2. Thermodynamics of clathrate hydrates 4
1.1.3. Molecular geometry and host-guest interactions 7
1.1.4. Kinetics of clathrate hydrate formation and dissociation 9
1.2. Clathrate hydrates for energy and environmental applications 15
1.2.1. Gas storage via clathrate hydrates 15
1.2.2. Gas separation via clathrate hydrates. 16
Part 1. Thermodynamic and Kinetic Studies of Clathrate Hydrates Systems for
Natural Gas Storage
Chapter 2. Hydrate-based methane storage with exploring cyclopentyl amine as a tunable
thermodynamic promoter 20
2.1. Introduction 20
2.2. Materials and Methods 23
2.2.1. Materials and sample preparations 23
2.2.2. Crystalline structure analysis using HRPD and PXRD 24
2.2.3. Characterization of the guest molecules using Raman and NMR spectroscopies 24
2.2.4. Phase equilibrium measurements 24
2.2.5. Eutectic concentration determination of T–xCPA phase diagram using DSC 25
2.3. Results and Discussion 26
2.3.1. Crystalline structural analysis of the CPA + CH4 hydrate 26
2.3.2. Thermodynamic stability analysis of the CPA + CH4 hydrate 29
2.3.3. Guest occupation analysis of the CPA + CH4 hydrate 31
2.4. Conclusions 40
Chapter 3. Hydrate-based multicomponent ternary natural gas storage with controlling
epoxycyclopentane concentrations for enhanced gas storage capacity 41
3.1. Introduction 41
3.2. Materials and Methods 45
3.2.1. Materials 45
3.2.2. Phase equilibrium measurements 45
3.2.3. Formation kinetics measurements and sample preparation 45
3.2.4. Crystalline structure identification 47
3.2.5. Cage occupation behavior characterizations 47
3.2.6. Calculation methods 48
3.3. Results and Discussion 50
3.3.1. Thermodynamic stability analysis of ECP + natural gas hydrates 50
3.3.2. Crystallographic structure analysis of ECP + natural gas hydrates 52
3.3.3. Formation kinetic measurement of ECP + natural gas hydrates 56
3.3.4. Guest distribution identification of ECP + natural gas hydrates 62
3.3.5. Time-dependent cage occupation behavior of ECP (6.5 mol%) + natural gas hydrate 66
3.4. Conclusions 74
Part 2. Thermodynamic and Kinetic Studies of Clathrate Hydrates Systems for
Storage and Separation from Hydrogen-Natural Gas Blends
Chapter 4. Promoting thermodynamic stability of hydrogen hydrates with gas-phase thermodynamic
promoters for hydrate-based hydrogen storage 76
4.1. Introduction 76
4.2. Materials and Methods 79
4.2.1. Materials and sample preparation 79
4.2.2. Phase equilibrium measurements 79
4.2.3. Crystalline structure analysis 80
4.2.4. Cage occupation behavior characterizations 80
4.2.5. Formation kinetics measurements 81
4.2.6. Calculation methods 81
4.3. Results and Discussion 83
4.3.1. Thermodynamic stability analysis of C2H6 + C3H8 + H2 hydrates 84
4.3.2. Crystallographic structure identification of C2H6 + C3H8 + H2 hydrates 86
4.3.3. Cage occupation behavior of C2H6 + C3H8 + H2 hydrates 88
4.3.4. Formation kinetic measurements of C2H6 + C3H8 + H2 hydrates 93
4.4. Conclusions 97
Chapter 5. Enhancing methane selectivity via tuned clathrate hydrate for methane enrichment from
hydrogen-natural gas blends 98
5.1. Introduction 98
5.2. Materials and Methods 101
5.2.1. Materials 101
5.2.2. Phase equilibrium measurements 101
5.2.3. Formation kinetics measurements and sample preparation 101
5.2.4. Crystalline structure analysis 102
5.2.5. Cage occupation behavior characterizations 103
5.2.6. Calculation methods 103
5.3. Results and Discussion 105
5.3.1. Thermodynamic stability analysis of THF, DIOX, and Dioxane hydrates 105
5.3.2. Formation kinetic measurements of THF, DIOX, and Dioxane hydrates 107
5.3.3. Crystallographic structure identification of THF, DIOX, and Dioxane hydrates 110
5.3.4. Cage occupation behavior of THF, DIOX, and Dioxane hydrates 112
5.3.5. Separation behavior for CH4 enrichment of THF, DIOX, and Dioxane hydrates 116
5.4. Conclusions 120
Chapter 6. Summary 121
Chapter 7. Conclusion 122
References 125
Curriculum Vitae 14
Investigating long-term training for remote sensing object detection
No full textRecently, numerous methods have achieved impressive performance in remote sensing object detection, relying on convolution or transformer architectures. Such detectors typically have a feature backbone to extract useful features from raw input images. A common practice in current detectors is initializing the backbone with pre-trained weights available online. Fine-tuning the backbone is typically required to generate features suitable for remote-sensing images. While the prolonged training could lead to over-fitting, hindering the extraction of basic visual features, it can enable models to gradually extract deeper insights and richer representations from remote sensing data. Striking a balance between these competing factors is critical for achieving optimal performance. In this study, we aim to investigate the performance and characteristics of remote sensing object detection models under very long training schedules, and propose a novel method named Dynamic Backbone Freezing (DBF) for feature backbone fine-tuning on remote sensing object detection under long-term training. Our method addresses the dilemma of whether the backbone should extract low-level generic features or possess specific knowledge of the remote sensing domain, by introducing a module called 'Freezing Scheduler' to manage the update of backbone features during long-term training dynamically. Extensive experiments on DOTA and DIOR-R show that our approach enables more accurate model learning while substantially reducing computational costs in long-term training. Besides, it can be seamlessly adopted without additional effort due to its straightforward design. The code is available at https://github.com/unique-chan/dbf.FALSEsciescopu
Improving esports viewing experience through hierarchical scene detection and tracking
No full textThe role of an observer in esports is to provide spectators with the most engaging scenes in real time. To automate this process, various research has been conducted. In this study, we utilize Vision Transformer (ViT)-based object detection to enhance the accuracy of automatic observers. However, while ViT-based detection more accurately identifies engaging game scenes, it often leads to frequent and abrupt scene changes, reducing viewer comfort. To address this issue, we propose a novel hierarchical structure that combines scene detection with scene tracking, maintaining high accuracy while ensuring smoother transitions between scenes. This approach also improves inference speed, as the tracking model is faster than the detection model. We computationally evaluated six observer models in terms of accuracy and camera stability, with our method demonstrating significantly more stable camera control. Additionally, user testing indicated a strong preference for our model over those without tracking. A video comparing our method to the state-of-the-art can be viewed at https://youtu.be/gWiU4GACZEg. © The Author(s) 2025.TRUEsciescopu
Etchant-Free Dry-Developable Extreme Ultraviolet Photoresist Materials Utilizing N-Heterocyclic Carbene–Metal Complexes
Full text linkExtreme ultraviolet (EUV) lithography has enabled significant reductions in device dimensions but is often limited by capillary force-driven pattern collapse in conventional wet processes. Recent dry-development approaches, while promising, frequently require toxic etchants or specialized equipment, limiting their broader applicability and highlighting the need for more sustainable, cost-effective alternatives. In this study, highly reactive, etchant-free dry-developable EUV photoresists using N-heterocyclic carbene (NHC)-based metal-ligand complexes, achieving half-saturation at EUV doses of 8.5 or 27 mJ cm−2, are synthesized. A simple thermal dry development process is employed, utilizing a standard furnace to remove unexposed areas of the photoresist, leading to 80nm resolution with line-edge roughness (LER) comparable to wet-developed patterns. Moreover, EUV-induced chemical reactions of the NHC metal-ligand complexes are investigated via EUV-photoelectron spectroscopy, near-edge X-ray absorption fine structures, X-ray photoelectron spectroscopy, and density functional theory. It is suggested that the high EUV sensitivity of the NHC metal-ligand complexes is attributed to branching polymerization reactions initiated by secondary electron and photoelectron generation. These EUV-sensitive, dry-developable NHC metal–organic photoresists offer a sustainable and economical alternative to conventional techniques, eliminating the need for toxic and corrosive etchants while achieving high-resolution nanopatterns through simple thermal treatment, thus advancing future nanofabrication technologies. © 2025 Wiley-VCH GmbH.TRUEsciescopu
Comparative Toxicity of Leachates from Menthol and Charcoal Cigarette Filter Microplastics on Caenorhabditis elegans
No full textLSE's Demand Bidding Strategy Considering Dynamic Generator Interactions and Transitional Market Characteristics in Korea Electricity Market
No full textKorea's electricity market is undergoing significant reforms, including the introduction of a real-time market and two-way bidding systems. As these changes unfold, Korea's single Load Serving Entity (LSE), which manages annual wholesale electricity purchases of approximately 60 trillion won, faces critical challenges in determining optimal bidding quantities in the wholesale electricity market across different seasons and time periods. This study employs reinforcement learning using Multi-Agent Proximal Policy Optimization (MAPPO) to develop bidding strategies that reflect the special characteristics of Korea's electricity market, while utilizing SHAP (SHapley Additive exPlanations) to analyze the importance of various features in the demand bidding process. The simulation results reveal that marginal generators typically offer 40% below their actual costs in their minimum generation capacity, suggesting potential market price reductions with generation-side competition. LSE's bidding strategies show consistent underbidding patterns across seasons, where analysis of best and worst episodes reveals significant strategic divergence during high-demand seasons characterized by temporally differentiated aggressive bidding, but minimal variance during low-demand periods suggesting limited strategic impact. The SHAP analysis identifies the strategy index as the most influential factor in bidding decisions, with demand error and LSE's action emerging as secondary factors during peak and off-peak hours, respectively. This study provides meaningful market outcomes through reinforcement learning simulation despite the absence of real-time market and generators’ offer data, while providing valuable insights for future market design implementations by incorporating Korea's special market characteristics.MasterAbstract ⅰ
List of tables ⅳ
List of figures ⅴ
Nomenclature ⅵ
Ⅰ. Introduction 1
Ⅱ. Special characteristics of the transitional Korean electricity market (TKEM) 6
2. 1. Two-settlement structure in Korean electricity market 6
2. 2. Two-way bidding pool (TWBP) in Korean electricity market 6
2. 2. 1. Limited price offer (LPO) in Korean electricity market
Ⅲ. Proposed bidding architecture for both LSE and generators 7
3. 1. Market assumption 8
3. 2. Market clearing algorithm 8
3. 2. 1. Day-ahead market 8
3. 2. 2. Real-time market 9
3. 3. Bidding problem in RL approach 10
3. 3. 1. Markov Decision Process (MDP) formulation 10
3. 3. 2. RL formulation of the bidding problem 10
3. 3. 3. Multi-Agent Proximal Policy Optimization algorithm (MAPPO) 13
3. 4. Shapley additive eXplanations (SHAP) for feature analysis 15
Ⅳ. Case studies 16
4. 1. Simulation data and implementation 16
4. 1. 1. Test system configuration and data preprocessing 16
4. 1. 2. Reinforcement learning parameter description 18
4. 2. Empirical Analysis and Discussion 19
4. 2. 1. Strategic offer patterns of marginal generators 19
4. 2. 2. Comparative analysis of strategic bidding behavior in LSE 21
4. 2. 3. Feature importance analysis for demand bidding using XAI 23
Ⅴ. Conclusion 25
References 26
Acknowledgement 2
Postfilter for multi-channel speech enhancement integrating spatial and spectro-temporal information
No full textSpeech enhancement aims to suppress the noise and improve the quality and intelli- gibility of degraded speech signals in a variety of acoustic environments. In contrast to single channel speech enhancement, which exploits the spectro-temporal information of the input signal to estimate clean speech, multichannel speech enhancement can further leverage spatial cues obtained from multiple microphones to achieve improved performance. There has been a growing demand for speech enhancement using micro- phone arrays in speech processing applications such as automatic speech recognition, mobile communications, meeting summarization and hearing aids. One of the popular configurations for multichannel speech enhancement is to apply a spatial filter exploit- ing spatial diversity of sound sources, and then apply a single-channel postfilter to the output of the spatial filter. One of the popular configurations in multichannel speech enhancement involves applying a spatial filter, such as the minimum variance distor- tionless response (MVDR) beamformer, that exploits the spatial diversity of sound sources, followed by a single-channel postfilter to further reduce residual noise. In this dissertation, we propose methods that incorporate multichannel information to construct postfilters aiming to effectively utilize spatial and spectro-temporal infor- mation. First, we propose a statistical model-based postfilter for dual-channel speech enhancement, combining a statistical model-based single channel noise power spectral density (PSD) estimator and a coherence-based dual channel noise PSD estimator with the a posteriori speech presence probability (SPP). We model the coherence-based a posteriori SPP as a simple function of the magnitude of coherence between two micro- phone signals and combine it with a single-channel SPP based on statistical models. The coherence-based estimator for the PSD of the noise remaining in the beamformer output in the presence of speech is derived using the pseudo-coherence considering the effect of the beamformers, which is used to construct the coherence-based noise PSD estimator. Then, the final noise PSD estimator is obtained by combining the coherence- based and statistical model-based noise PSD estimators with the proposed SPP. The spectral gain function is also modified, incorporating the proposed SPP. Experimental results demonstrate that the proposed method led to more accurate noise PSD estima- tion and perceptual evaluation of speech quality (PESQ) scores in various diffuse noise environments, and did not degrade the speech quality under the presence of directional interference, although the proposed method utilizes the coherence information. Secondly, we propose a deep neural network (DNN)-based postfitler for multichan- nel speech enhancement, integrating DNN-based parameter estimation for multichannel speech enhancement utilizing parameters estimated from the beamforming stage in the parameter estimation for the postfilter. Specifically, the a posteriori SPP, the speech PSD, and the noise PSD in estimated in the beamforming stage are utilized to im- prove the parameter estimation in the postfiltering stage. We also adopt the dual-path conformer structure with an encoder and decoders to enhance the performance. Ex- perimental results show that the proposed method marked the best wideband PESQ scores on the CHiME-4 dataset among all compared methods. ©2025 Sein Cheong ALL RIGHTS RESERVED – iii –|음성 향상 기술은 다양한 음향 환경에서 잡음을 억제하고 음성 신호의 품질과 명료 성을 개선하는 기술로, 최근 음성 인식, 모바일 통신 및 보청기 등 실제 응용 분야에서 수요가 지속적으로 증가하고 있다. 입력 신호의 스펙트럼-시간 정보만을 활용하는 단일 채널음성향상기법과달리,다채널음성향상기법은여러마이크로폰에서얻어진공간 정보를 추가적으로 활용함으로써 더욱 우수한 성능을 얻을 수 있다. 다채널 음성 향상의 대표적 구성 방법 중 하나는 최소 분산 비왜곡 응답(minimum-variance distortionless response, MVDR) 빔포머(beamformer)와 같이 음원의 공간 정보를 활용하는 공간 필터 (spatial filter)를 먼저 적용하고, 공간 필터의 출력에 단일 채널 후처리 필터(postfitler) 를 추가로 적용하여 잔여 잡음을 제거하는 방식이다. 본 학위 논문에서는 공간 정보와 시간-주파수 정보를 효과적으로 결합한 다채널 음 성 향상 후처리 기법들을 제안하였다. 먼저 두 채널 음성 향상을 위한 통계 모델 기반 후처리 기법을 제안하였다. 본 방법은 통계 모델 기반의 단일 채널 잡음 전력 스펙트럼 밀도(power spectrum density, PSD) 추정기와 두 마이크로폰 신호의 일관성(coherence) 을 이용한 두 채널 잡음 PSD 추정기를 결합한다. 두 마이크로폰 신호 간의 coherence 의 크기를 활용하여 coherence기반 사후 음성 존재 확률(a posteriori speech presence probability, SPP)을 모델링하고 이를 통계 모델 기반의 단일 채널 a posteriori SPP를 결합하였다. 그 다음, 음성 존재 불확실성을 반영하여 coherence기반의 두 채널 잡음 PSD 추정기를 유도하고, 이를 통계 모델 기반 잡음 PSD 추정기와 a posteriori SPP를 활용하여 결합한다. 최종적으로 결합된 잡음 PSD 추정치와 a posteriori SPP를 이용하 여최적변형로그스펙트럼진폭(optimally modified log-spectral amplitude, OM-LSA) 이득 함수를 계산하여 후처리 필터로 사용하였다. 이를 통해 빔포머 출력 신호에 남아있 는 잡음의 PSD 추정 정확도를 높였으며, 다양한 확산 잡음(diffuse noise) 환경에서 음성 품질(PESQ 점수)을 향상시켰다. 두 번째로는 심층 신경망(deep neural network, DNN) 기반 다채널 음성 향상 후처 리 기법을 제안하였다. 본 방법에서는 빔포밍 단계에서 추정된 음성 PSD, 잡음 PSD, a posteriori SPP 등의 파라미터를 후처리 단계의 파라미터 추정 과정에 통합하여 공간 정보와 스펙트럼-시간 정보를 종합적으로 활용하였다. 추가로, dual-path conformer 네 트워크 구조를 도입하여 파라미터 추정의 정확도와 계산 효율성을 동시에 향상시켰다. 실험 결과, CHiME-4 데이터셋에서 제안한 방법이 기존 방법 대비 가장 높은 음성 품질 (PESQ 점수)을 나타냈다. 결론적으로, 본 논문에서 제안한 기법들은 단일 채널 및 다채널 음성 향상 기술과 딥러닝 접근법을 효과적으로 결합하여 낮은 계산 복잡도에서 강건한 성능을 달성하였 다. 이에 따라 제안된 방법들은 음성 인식, 모바일 통신, 보청기 등의 실제 음성 처리 분야에서 실질적으로 활용될 수 있을 것으로 기대된다. ©2025 정 세 인 ALL RIGHTS RESERVEDDoctorAbstract (English) i
Abstract (Korean) iv
List of Contents vi
List of Tables viii
List of Figures ix
1 Introduction 1
1.1 Introduction 1
1.2 Outline of the Dissertation 3
2 Postfilter for Multichannel Speech Enhancement Using Coherence
and Statistical Model-based Noise Estimation 5
2.1 Introduction 5
2.2 System Overview and Review of SPP-Based Noise Estimation 8
2.2.1 Problem Formulation and System Overview 8
2.2.2 Single Channel Noise PSD Estimator Based on Speech Presence
Probability 10
2.3 Postfilter for dual channel speech enhancement utilizing noise estimation
based on coherence and statistical model 12
2.3.1 Modeling of a Posteriori SPP Based on Coherence 14
2.3.2 Proposed Dual channel Noise PSD Estimator Based on Coherence 16
2.3.3 Combining Noise PSD Estimates and Gain Calculation 22
2.4 Experimental and Results 23
2.4.1 Experimental Configurations 23
2.4.2 Experimental Results 27
3 Integrated DNN-based Parameter Estimation for Multichannel Speech
Enhancement 39
3.1 Introduction 39
3.2 Review of DNN-based Parameter Estimation 41
3.2.1 MVDR Beamforming and Postfiltering 41
– vi –
3.2.2 Parameter Estimation for Beamforming and Postfiltering 44
3.3 Integrated DNN-based Parameter Estimation for Multichannel Speech
Enhancement 46
3.3.1 Architecture of the DNNs 48
3.3.2 Postfiltering with Integrated Estimation of PSDs 49
3.4 Experimental and Results 52
3.4.1 Experimental Configurations 52
3.4.2 Experimental Results 53
4 Conclusions 55
References 58
– vii