High-resolution Algorithm for Wind Data near Nuclear Power Plant : Integrated Approach of Deep Learning and Singular Value Decomposition

이다경 서울대학교 대학원 2024 국내석사

본 연구는 원전 사고 시 발생할 수 있는 방사성 물질의 확산 및 분산 평가를 위해 상세한 기상 자료의 필요성을 다룬다. 한국 내 원전이 주로 동해와 서해 연안에 위치하고 있는 점을 고려할 때, 복잡한 지형과 변동성이 큰 날씨 패턴을 반영하는 정밀한 대기 확산 모델과 고해상도 기후 데이터의 중요성이 강조된다. NCEP, ECMWF와 같은 기존의 전 지구 기상 데이터는 해상도의 한계로 인해 한반도의 특정 지형 및 기상 패턴을 충분히 반영하지 못하며, 이로 인한 방사성 물질 확산 모델링의 예측 오류를 초래한다. 고해상도 기후 데이터를 얻기 위한 다양한 방법이 연구되고 있지만, 이러한 접근 방식들은 상당한 계산 비용을 동반하거나 여러 가지 신뢰성 문제를 내포하고 있다. 따라서 기존 방법들의 한계를 극복하기 위한 새로운 고해상화 방법이 필요하다. 또한, 고해상도 데이터로의 데이터 향상 과정은 데이터양을 기하급수적으로 증가시킬 가능성이 있다. 기후 데이터와같이 다양한 변수를 포함하는 경우 이러한 데이터 해상도의 증가는 기후 데이터의 비효율적인 저장 및 이용 모두를 초래할 수 있다. 이러한 문제들을 해결하기 위해, 본 연구는 물리 기반 WRF 모델링과 최신 딥 러닝 기술을 결합한 새로운 프레임워크를 개발하고 검증한다. 이 혁신적인 접근법은 전통적인 물리 모델링의 높은 계산 비용과 복잡성과 단순 보간 방법이 내재한 보간 오류를 줄일 수 있다. 또한, 고해상도 기후 데이터의 증가된 데이터 볼륨 문제를 해결하기 위한 특잇값 분해(SVD)를 데이터 압축에 효과적으로 사용한다. 풍속과 풍향(U 및 V 변수)에 중점을 두고, 고리 원자력 발전소 인근(반경 20km 이내)에서 바람 데이터의 고해상화를 수행한다. 본 연구의 결과는 다음과 같다. 딥 러닝 모델 Uformer의 Deblurring 기술을 활용하여 바람 데이터의 수평 해상도를 1.5km에서 300m로 5배 향상했다. 또한, 특잇값 분해(SVD)를 통해 고해상화 기후 데이터의 용량을 약 30% 압축하였다. 기후 데이터를 고해상화 하는 전통적인 방법의 하나인 WRF를 이용했을 때보다 딥 러닝인 Uformer를 이용하여 고해상화 결과를 출력했을 때 최소 97%의 시간 향상이 있음을 보였다. 또한, 라그랑지안 입자 확산 모델(LPDM)에 데이터 압축을 위해 특잇값 분해(SVD)를 적용하였는데, 이를 통해 원 데이터와의 1.2E-06의 작은 오차로 전체 용량의 30%를 절약할 수 있었다. 그러나, 특잇값 분해(SVD)를 적용하면서 데이터 로딩 방법에 변화가 발생하게 되어 3.41%의 적은 시간 지연이 발생함을 확인하였다. 제안된 방법론은 더 빠르고 효율적으로 고해상도 기후 정보를 제공하며, 데이터 압축을 통해 사용 편의성을 개선하는 것을 목표로 한다. 기존 기상 데이터 고해상화 방법의 한계를 극복하고, 한국 원전 지역의 복잡하고 다변적인 기후 맥락에서 방사성 물질 확산 예측의 정확도를 향상시킬 것으로 기대된다. 주요어 : 딥 러닝, 바람 데이터, 고해상화, 기상 연구 및 예측 모델(WRF), 특잇값 분해(SVD) 학번 : 2022-26504 This study deals with the need for detailed meteorological data to evaluate the diffusion and dispersion of radioactive materials that may occur during a nuclear accident. Considering that nuclear power plants in Korea are mainly located in the East Sea and West Sea coasts, the importance of precise atmospheric diffusion models and high-resolution climate data reflecting complex topography and ever-changing weather patterns is emphasized. Existing global meteorological data such as NCEP and ECMWF do not sufficiently reflect specific topography and weather patterns on the Korean Peninsula due to the limitations of resolution, resulting in predictive errors in radioactive material diffusion modeling. Various methods are being researched to obtain high-resolution climate data, but these approaches often come with significant computational costs and reliability issues. Therefore, there is a need for new high-resolution techniques to overcome the limitations of existing methods. Additionally, enhancing data resolution can lead to a geometric increase in data volume, which can be detrimental to both storage and utilization, especially for climate data that encompass numerous variables. To address this challenge, this study develops and validates a new framework that combines physics-based WRF modeling with state-of-the-art deep learning techniques. This innovative approach aims to reduce the high computational cost and complexity associated with traditional physics modeling and mitigate interpolation errors inherent in simple interpolation methods. Additionally, it effectively utilizes Singular Value Decomposition (SVD) for data compression to address the increased data volume of high-resolution climate data. The focus of this research is on wind speed and wind direction (U and V variables), and it performs high-resolution data generation for wind data near the nuclear power plant vicinity (within a 20 km radius). The results of this study are as follows. By utilizing the Deblurring technique of the deep learning model Uformer, the horizontal resolution of wind data has been improved from 1.5km to 300m, a five-fold enhancement. Additionally, through Singular Value Decomposition (SVD), the capacity of high-resolution climate data has been compressed by approximately 30%. When compared to using the traditional WRF method for high-resolution data generation, using the deep learning Uformer model resulted in a minimum 97% improvement in processing time. Furthermore, when applying SVD for data compression in Lagrangian Particle Dispersion Model (LPDM), it was possible to save 30% of the total capacity with a negligible error of 1.2E-06 compared to the original data. However, the application of SVD introduced changes in the data loading process, leading to a minor time delay of 3.41%. The proposed methodology aims to provide high-resolution climate information more quickly and efficiently while improving user convenience through data compression. Its goal is to overcome the limitations of existing methods for enhancing weather data resolution and enhance the accuracy of predicting the dispersion of radioactive materials in the complex and diverse climate context of South Korean nuclear power plant regions. Keywords : Deep Learning, Wind data, High-resolution, Weather Research & Forecasting Model (WRF), Singular Value Decomposition(SVD) Student Number : 2022-26504

Depthwise Channel Attention Network (DWCAN): An Efficient and Lightweight Model for Single Image Super-Resolution (SISR) : An Efficient and Lightweight Model for Single Image Super-Resolution (SISR)

YASIR MUHAMMAD 가천대학교 글로벌캠퍼스 일반대학원 2024 국내석사

Single image super-resolution (SISR) is a popular technique in image processing, often using deep learning models such as convolutional neural networks (CNNs). However, these models can be complex and resource-intensive, making them challenging to use on devices with limited resources. Researchers have developed resource- constrained models that use fewer parameters and computations to address this issue. Examples include the information distillation network (IDN), the information multi-distillation network (IMDN), and the residual feature distillation network (RFDN). However, further optimization is necessary to improve the efficiency and compactness of these models. We present a new lightweight super-resolution (SR) model based on the RFDN architecture, which won the AIM2020 and NTIRE2022 SR competitions. Our proposed Depthwise Channel Attention Network (DWCAN) makes several changes to RFDN. It substitutes the main residual feature distillation block (RFDB) with a depthwise channel attention block (DWCAB). DWCAN also adds a shallow residual block (SRB) incorporating depthwise separable convolution (DW) and a channel attention (CA) block. We aim to reduce model complexity, computations, inference speed, and memory usage while retaining or enhancing the peak signal-to-noise ratio (PSNR). The experimental results demonstrate the effectiveness of the proposed model. By applying our modifications, we achieve a notable reduction in model complexity, leading to an improved PSNR of 29.07 decibels (dB), up from RFDN's 29.04 dB on a diverse 2K resolution (DIV2K) dataset. This underscores the potential of our lightweight model to balance computational efficiency and SR quality. Our technique lowers model complexity, improving image quality with a PSNR score of 29.07 dB on the 2K resolution dataset (DIV2K), exceeding RFDN's 29.04 dB. This shows our model's potential for balancing performance and efficiency.

不實債權 整理를 爲한 法的基盤에 관한 硏究 : 國際 標準化를 中心으로

정재룡 서울大學校 大學院 2004 국내박사

우리나라를 포함하여 1997년 금융 및 외환위기로 인하여 경제위기를 겪은 동아시아 국가 등 대부분의 신흥시장 경제 국가와 채제전환국 및 일본의 부실채권시장의 공통적인 특징은, 경제위기로 인해 발생한 부실채권을 처리하기 위해 정부가 설립한 공적기관의 주도 하에 부실채권정리작업이 추진되어 왔다는 점이다. 이들 국가에 있어서 不實債權의 조기정리를 위한 법적 제도적 인프라가 미약하여 민간부문이 부실채권정리에 효과적으로 대응하지 못하였기 때문이다. 본 논문에서는 주로 우리나라, 일본 등 금융시스템의 비효율성으로 경제위기를 겪었거나 현재 겪고 있는 나라와 중국과 같은 체제전환국 및 말레이시아와 같은 신흥시장경제국가들의 현실적인 부실채권정리 법제 하에서, 부실채권정리 활성화를 위해서는 國際標準化가 필요하다는 當爲性을 규명하고 이를 위한 구체적 방안을 연구하고자 한다. 국제표준화란 일반적으로 여러 나라 사이에 통용되거나, 또는 여러 나라를 포괄하여 마땅히 따르고 지켜야 할 본보기나 법식, 또는 제도를 의미한다. 국제표준화의 명제를 부실채권에 관한 법제를 국제적 규범으로 만드는 노력을 경주하여 부실채권정리 시장 내에서 공정거래가 이루어지고 부실채권이 효율적으로 정리 되도록 함으로써 부실채권정리의 국제협력을 통하여 세계경제의 활성화를 도모하자는 차원에서 접근한다. 경제가 글로벌화해지면서 각국의 경제에 있어 금융도 글로벌화되어 가고 있다. 이에 따라 어떤 나라의 금융의 부실은 이웃나라의 부실로 이어지는 同時化現像이 나타나고 있다. 이러한 점을 감안할 때 세계경제 전체의 금융구조의 건실화를 위한 방안의 모색은 불가피한 것이다. 이를 위해서는 부실채권정리가 현안이 되어 있는 개별 국가가 자주적으로 부실채권정리문제를 이미 경험한 선진제국의 제도나 최근 모범적인 정리실적을 보인 국가의 제도를 도입하는 경우도 있을 수 있겠으나, 이 보다는 국제적인 협력을 통해 이러한 제도의 도입을 여러 나라에 적극 확산함으로써 부실채권시장을 육성하는 것이 각국의 경제, 나아가서는 세계경제의 안정적인 성장의 밑거름이 된다고 생각한다. 부실채권정리관련 법제의 표준화는 다음의 두 가지 방향에서 모색할 수 있을 것이다. 첫째는 자국의 부실채권정리에 있어서 성공적 성과를 거둔 국가의 제도와 경험을 공유하도록 하는 방법이다. 성공한 사례로서 예시된 국가의 제도를 국제모델화하여 러시아, 중국, 인도네시아 등과 같은 體制轉換國이나 신흥경제국가에 전파하도록 하는 것이다. 둘째는 국제금융 및 자본시장에 있어 부실채권에 대한 투자여건을 개선하는 방법이다. 국제표준화는 지역적으로 이루어지거나 또는 범세계적으로 이루어지는 경우로 나누어 볼 수 있으나 대부분의 경우 일부 지역 또는 국가에서 먼저 형성된 국제규범이 점차 다른 국가나 지역으로 전파되는 경로를 따르게 된다. 또한 각국의 고유한 문화나 법체계의 특성을 감안하여 논의가 전개되어야 하며, 실효성과 법적 안정성으로 대표되는 상충하는 가치에 대한 비교형량의 과정이 뒤따르게 될 것이다. 구체적인 추진방안에 관해서는 韓國資産管理公社의 경험이 유용할 것이다. 즉 한국자산관리공사는 상대방 국가의 공적기구와 상호 제도연구 및 통일화를 위한 노력을 경주한다는 양해각서를 교환하고, 나아가 국제적인 협력체계를 구축하기 위한 국제포럼의 개최를 주도적으로 수행하였다. 이와 같은 한국자산관리공사의 역할은 부실채권 정리에 관한 국제적인 룰을 형성하는데 있어서 긍정적이었던 것으로 평가할 수 있다. 표준화에 관하여 모델의 구체적인 내용은 어느 한 나라의 제도에 전적으로 의존하는 것보다는, 각국의 제도를 다각적으로 연구하여 하나의 모자이크법을 구성하는 것이 바람직할 것이다. 실제 모델법령에 수용하여 각국에 전파할 가치가 있는 것으로 판단되는 제도는 그 동안 제도운영의 결과가 비교적 성공적인 국가의 것들이다. 이러한 점에서 한국, 말레이시아 또는 미국의 제도는 특히 본받을 가치가 있는 것으로 생각된다. 이들 국가의 부실채권정리제도의 공통점은 정부 또는 공적기구가 강력한 권한을 가지고 공적자금을 활용하여 부실채권을 신속하게 정리하고자 하였다는 점이다. 이러한 접근방법은 부실채권시장의 형성·발전단계로 보아 제2단계 내지 제3단계에 해당하는 것으로 완숙한 시장의 기능에 의존할 수 없는 특수한 상황을 전제로 하는 경우일 것이다. 이상의 논의에 기초하여 볼 때, 모델법령의 내용에는 不實債權의 基準과 範圍에 관한 사항, 부실채권정리 재원의 마련을 위한 제도 및 각종 절차의 單純化 및 統一化에 관한 사항들이 반영되어야 할 것이다. This paper investigates regulatory policies in the resolution of the NPLs. The researcher also studies about legal approaches in the resolution of NPLs in order to make the NPLs resolution process more effective and to vitalize international economy in the long term. Also, this paper suggests that international standardization of relevant legislation is an important way to resolve the NPLs more effectively. It is important to improve the efficiency in the financial sector of an economy so that the market can function effectively. Efficient asset management is one of the most important factors in order to achieve the stability in the financial sector. Too much NPLs held by the financial sector not only leads to the bankruptcy of individual financial institutions but also shrinks the overall business activities in the financial sector. Furthermore, if the situation gets worse, it would weaken the competitiveness of the financial businesses, and eventually lead to a credit crisis at the national level. It also can hinder the solidity of financial infrastructure. Considering the trend of internationalization and liberalization in the financial market, the resolution of NPLs is very important for the sound management of banks and the stable management of the national economy. In this regard, in emerging market economies such as Korea and Malaysia including South East Asian countries, the successful disposal of NPLs is very important for sustainable economic growth in the future. In order to resolve NPLs effectively, private NPL resolution agencies need to play an active role in the market. NPLs are relatively more difficult to be evaluated than normal performing assets, and they need to be handled in sophisticated ways. In nature, NPLs are difficult to be sold in the normal asset transaction market. Considering these characteristics of NPLs, many countries have established special institutions for the resolution of NPLs. In countries like the U.S.A., Germany, U.K., where financial markets are well developed, there are plenty of private asset management companies and investment companies. In some countries other than those mentioned above, governments are establishing public companies for the resolution. NPLs market constitutes of three entities: the provider who helps to dispose NPLs; the broker-dealer who wants to make the transaction; and investors either institutional or individual as the final consumer. Countries which experienced the financial crisis in 1997, including Korea and other East Asian countries, have established government-driven public institutions to resolve NPLs. In those some other countries mentioned above, it was proved in general that their private sectors are relatively inefficient in resolving NPLs, and the inefficiency originates from the weakness of their legal infrastructure to a significant degree. Thus, the resolution of NPLs are planned and executed by public institutions at the government level. Government-led resolution of NPLs has some advantages in terms of transparency, expertise, and independency in the process compared to that managed by private companies. However, this also means that there could be difficulties in evaluating the fair market value of NPLs. Investors in the international financial market consider many things before making decisions to invest in the NPLs of foreign countries. They are the potential profitability and the stability of the market, etc. The legal infrastructure as well as tax benefits are also considered seriously. In addition, the following aspects are important; whether investors can manage the liquidity and risks of their investment in the local securities market; whether financial institutions are well equipped with techniques to increase the liquidity of acquisition asset; whether there are lots of human capital with expertise; whether the system for information transmission and management is effective; whether regulatory systems for the corporate split-up, mergers & acquisition, reorganization, and liquidation as well as those for the protection of investors, including regulatory alleviation and corporate disclosure are well built. For the newly emerging market economies such as Korea, Japan, and South Eastern countries to attract investors effectively, more efforts need to be made to establish such infrastructures. This paper researches about regulatory policies for the resolution of the NPLs, of several countries especially United States, Korea and Malaysia, which at some time in their history experienced economic crisis due to ineffective financial system. Also, this paper takes into account the fact that newly emerging economies such as China and Eastern European countries need to develop practical regulatory policies to resolve NPLs. The researcher tries to find solutions for the betterment of the process resolving NPLs in legal a perspective. Also, this paper points out that international standardization is necessary for the development of the systems for the resolution of NPLs. International standards could be defined as the exemplary model or system to be applied in the international level. In this paper the researcher approaches the issue of resolving NPLs in terms of enhancing fair trade and inducing decent growth in the international economy. Effective resolution of NPLs through global cooperation would lead to the stabilization of financial systems and would eventually lead to the international economic growth. Financial system in a society is often compared to the blood circulation in human body. Financial insolvency can result in the economic crunch in a society. It can be compared to the phenomenon that hardening of arteries cause the human body to collapse either gradually or all of a sudden. In general, the phenomenon that the barriers to the flow of the economic transactions are abolished leads to the financial globalization. In accordance, the financial crisis in a country consequently can lead to the financial problem in the countries which have economic relations with the country in trouble. Considering the ever deepening of globalization in the financial transactions, it would be the most imperative to search for effective ways for the stabilization of the international financial system. As a way for that purpose, solidification of the system to resolve NPLs especially through international cooperation would be very helpful. This will also contribute to the sustainable growth of world economy as well as to the enhancement of economic stability of each country. On the other hand the financial system of each country which includes the system to resolve NPLs is inter-related with those of other countries. So building up the effective system to resolve NPLs would not be a simple agenda restricted to in a few countries, considering the inter-related nature of financial system at the global level. This paper explores systematic ways to improve the problem of NPLs with an emphasis on regulation. For example, if issues like adjustment of double taxation and the free movement of international capital could be improved, NPLs could be resolved more efficiently worldwide. Two approaches could be taken in the discussion of the internationalization of legal system with respect to NPLs resolution. First, it would be helpful to share individual nation's regulatory policies and experiences in the resolution of NPLs. Successful examples in the resolution in some countries such as Korea and Malaysia could be followed by other countries such as Russia, China, and South Eastern countries, where the legislative efforts with respect to NPLs resolution are rather limited. Second, it is important to improve the attractiveness of NPLs as investable assets from the perspective of investors in the international and financial markets. NPLs may be considered as an attractive investment by domestic and international investors if they can show good points in the aspects such as profitability, stability and liquidity. In reality, foreign investors have experienced lots of difficulties in the foreign markets since each country has its own special system and criteria for the resolution of NPLs, especially in terms of government regulation on the capital flow. More specifically, in countries where NPLs accumulation is huge that they cannot resolve them effectively with domestic capitals, the strong government regulation has been a pivotal problem. In such countries, it would be helpful for the government to revise the system so that foreign investors can invest their capital actively in the NPLs market without any doubt that they are not compete under fair condition. For foreign investors, reasonable regulatory system is one of the very important factors for the decision of investment. In nature, each NPLs market has its own particularities. Thus, it is necessary to consider the history and circumstances of each country when the government establishes the legal system for the resolution of NPLs. In general financial systems of a country are affected by its particular circumstances. Therefore, it is important to consider individual nation's specialties even if the internationalization of NPLs-related legislation is sought after. In this regard, we might be able to learn lessons from the GATS financial clauses of the WTO. The GATS financial clauses prepared during UR negotiations point out that the opening of the domestic market in the financial service sector needs to be based on individual nation's measures and systems. It clearly indicates that each country needs to recognize the differences in other nations' financial regulatory system. International cooperation is essential in the process of harmonizing different financial systems. In order to achieve this goal, it is necessary to recognize other nation's governmental regulations and to consult over the differences with each other. Harmonization of different financial systems which actually resluts in the standardization can be facilitated through international cooperation for mutual exchange of information, sharing common regulation, or establishment of new international organization for standardization and fund raising. In this regard, this paper explores the possibility of enhancement of international standardization in the resolution of NPLs. This paper also emphasizes that it is important for individual nations to recognize other countries' legislative history and economic circumstances in order to facilitate international standardization. This paper explores the possibilities of international standardization by classifying participants in the NPLs resolution market. The participants in NPLs market can be classified into provider, consumer, and broker-dealer. The providers are public organizations which purchase NPLs from various financial institutions and then resolve through direct selling. Investors as consumers own capital resources. Usually, these entities include asset securitization company, REITs, Corporate Restructuring Company(CRC), and Corporate Restructuring Vehicle(CRV). Dealers and brokers include security companies taking care of security acquisition and sale, consulting firms and accounting firms providing financial advice, credit rating agencies, and law firms as legal advisor. NPLs are provided by the financial institutions which are creditors to the debtors in trouble. Theoretically creditors can wait until the maturity date to exercise relevant legal interests, or they can go through certain procedures of loan securing by attachment. However, through a well developed NPLs resolution system, creditors can give debtors the opportunity to restructure companies. The creditors can also have opportunity to liquidate capital resources. It would also mean a new investment opportunity for general investors. What's important in this process is whether the debtor can actually pay for the original debt. Financial institutions first evaluate whether the loan is insolvent or not, and then they take care of the loans in a number of ways. Public resolution agencies undertake NPLs from financial institutions and then resolve them. In this process, financial institutions and public organizations follow domestic standards in evaluating the nature of NPLs and in resolving them. Financial institutions and public resolution agencies usually dispose the NPLs by issuing ABS, international bidding, individual disposition, or establishing joint venture company (AMC). While going through these processes, investors naturally weigh the future cash flow. Matters of weight and importance include credit line or credit enhancement through warranty, future expected exchange rate, transferability of capital gain, and statutory stability in the resolution of legal conflict. For a reasonable price to be established in the market, appropriate regulatory systems need to be in place. First, it is necessary to maintain effective relationship among banks, public organization, and the government. In order to achieve the effective relationship among different entities, it is necessary to ensure autonomous management in financial institutions as well as efficiency and transparency in public organizations. Next, NPLs as goods in the market need standardized criteria for accurate evaluation. It would be helpful to have consolidation of international standard and the establishment of credit screening(review) system. Furthermore, accounting standards in corporate enterprises and financial institutions need to be approximated to the international accounting principles. The disposal of NPLs could be pursued in the following area; issuing ABS, international bidding, individual disposition, and establishing joint venture company. The resolution of NPLs would be more active if NPLs are in the form of ABS. For issuing domestic ABS, it is required to go through the listing and disclosure process. Also, ABS laws regulate the conditions and procedures. However, in Korea, issuing ABS to resolve NPLs was not proved to be effective in attracting foreign capital. In order to attract international investors, internationalization in the credit rating system would be a key factor. Legislative efforts need to be placed in order for the ABS companies to ensure their rights for the credit. In the long term, securitized NPLs could be listed in the security market, if the international credibility of listed information increases. In this case, the value of NPLs as an object for investment would become higher. International bidding method has been used in Korea, U.S.A., Thailand, and Indonesia. Considering the competitive nature of bidding, this method could maximize the value of NPLs. Also, it is an effective way to let foreign investors know the value of NPLs. The process of international bidding needs to be standardized to minimize the purchase deposit and the cost associated with the process. This would results in the magnification of the subject assets and the mitigation of oligopoly behavior of major investment companies. Individual disposition includes Merger and Acquisition, individual Corporate Loan Sale, etc. This is a method often used in the U.S.A., Thailand, and Indonesia. Individual disposition is influenced by conservative legislation in each nation. As a part of NPLs, real estate properties or collateral of insolvent companies could be resolved through independent auction or court auction. Foreign investors could participate in the auction. Increasingly, there is a trend in many countries to allow for foreigners to acquire real estate properties. Regulations need to be made clear for the realty and credit related matters. However, regulations about real estate properties is usually based on individual nation's legislation and tradition. Thus, it would not be easy to pursue international standardization in this area. However, it is possible to explore ways to internationally circulate asset securities with the guarantee of NPLs and real estate properties. This is a method to establish asset management companies, corporate restructuring companies, or corporate restructuring vehicles as a joint effort between organizations which own NPLs and international investment organizations. When a foreign investment company establishes an AMC with an organization which owns NPLs, the price of NPLs could be increased, since the foreign investment company operates the capital market without intervention from the government. Foreign companies with capitals could participate in the NPL market, and they would expect to see the internationalization in issuing ABS. Corporate restructuring companies specialize in corporate undertaking, normalization of management, as well as direct selling. When CRCs are established by a joint effort between companies retaining NPLs and foreign investment companies, there are a number of advantages of using the know-hows and the capital of foreign investment companies. This case also requires the internationalization of related legislation. Investors as consumers of NPLs exhibit variety in terms of the transaction method and level. Realistically, the ultimate capital sources are domestic/international fund or financial institutions. Providers use either direct transaction with them or use intermediate buyers to make deals. Intermediate buyers are usually private enterprises who work within the market mechanism. In Korea, they are asset securitization companies, REITs, Corporate Restructuring Companies(CRC), and Corporate Restructuring Vehicle(CRV). Because of the high risk factors involved in the intermediate buyer structure, the intermediaries are entitled to have privileges in the contribution and assumption of liability. There are not many international players in the market, since investing in Non-Performing enterprises entails high risk factors. However, considering the increasement of vulture fund in the international capital market as well as in the junk bond market, a similar trend in the corporate restructuring market could be expected. When the final consumer is the foreign capital, the mitigation of various regulation becomes an important issue. For example, it would be attractive for foreign investors to see less regulation on foreign exchange trade, as well as a better system to ensure the stability of foreign exchange market. On the other hand, NPLs market has some speculative factors. When international Hedge Fund (Hot money) comes in, the market could be dominated by the Hedge Fund. In order to protect the market, policies need to be made. Consumers are expected to participate in the market under the principle of fair competition. Fair trade laws could provide regulation against mutual prearrangement or the spread of unfair information. Basically, efforts need to be made to maximize investors' participation through securitization, rather than through individual disposition. In general, brokers/dealers are not the direct participants in the transaction, but they provide with services to close the deal. In the NPLs resolution market, brokers and dealers work as intermediaries between the provider and the consumer. For providers, they exhume consumers with enough capital. For consumers, they provide with information about the goods. In Korea, security companies, credit evaluation companies, consulting firms, and accounting firms offer this service. These institutions provides with expertise to the market and they follow restricted laws in each nation. To protect investors in individual countries, operation process regulations and ethical regulations are written. Under certain circumstances, they have legal responsibilities. International standardization in this involves improvements in regulatory systems. The evaluation of NPLs is an important process from the perspective of investors. There are many risk factors due to the high uncertainty of NPLs. Insurance is one way to prepare for the possible dispute in the future, even though the insurance premium could reduce the earning rate. When not insured, it would be more difficult to prove a causal relationship in the case of a claim for damages. In pursuing the international(Global) standardization of relevant legislation, it is necessary to determine who would be the entity in charge. Generally, the international(Global) standardization in relevant legislation is processed either by a joint effort of a number of nations, or through an existing international organization. As a practical approach, the international standardization could be pursued and established in certain parts of the world and it could be transferred to a larger area to include more countries. The nature of international standardization would be different, depending on individual nation's circumstances in which the process occurs. For example, the process would results in a very different economic regulations in different countries, based on the cultural and social-economic contexts. Even though the principle of the financial market would expect to see the maximized standardization of financial transaction in as many countries as possible, the circumstances in which each country's financial market operates are important factors to consider in reality. Therefore, the process of pursuing international standardization needs to take a number of different and cautious approaches. First of all, it would help to examine the international standardization in the field of general economic regulations in exploring possibilities of the international standardization in resolving NPLs. The international standardization of economic regulations can be categorized into two: 1) the establishment and transfer of the standards; and 2) the bilateral agreement between two parties. Again, The autonomous acceptance of the international standard could be accomplished by the following four ways: 1) Governments could cooperate to standardize the procedures; 2) Governmental agencies of private institutions residing in different countries could standardize the procedures; 3) A model of agreement could come to a conclusion among different countries; and 4) A model of domestic regulations could be transferred to other nations. Also, more directly, international standards could be set up through both parties agreement or multi-parties agreement. Governmental agencies of private institutions residing in different countries could standardize the procedures. Examples include: Capital Adequacy ratio of the BIS, Bank of International Settlement; and the Labor Standard of OECD. The Capital Adequacy ratio of the BIS(Bank of International Settlement) is an agreement between the Presidents of central banks in Western developed countries. The ratio has been set for the standardization of the process in financial supervision. It entails mandatory power. A model of agreement could be proposed among different countries. An example would be the model of tax agreement, including the UN model and OECD model. The U.S.A. has its own model of agreement. The U.S.A. model exerts strong influence on the counterparts in terms of the mandatory power. In the case of the UN model, the perspectives of the developing countries are well represented. The OECD model generally represents the perspectives of developed countries. With the trend of valuing free market system, the OECD model is observed in these days. The model of tax agreement plays an important role in the international business environment. The Bankruptcy Act in association with the corporate restructuring has shown the trend of international standardization. In a broader concept, the resolution of NPLs includes the liquidation of insolvent companies and corporate restructuring. The regulations on the restructuring of insolvent corporations in Korea are the bankruptcy-related 3 acts and the Corporate Restructuring Investment Company Act. There also is an international trend of integrating many different bankuptcy-related acts into one. First of all, the UNCITRAL(United Nations Commission on International Trade Law) proposes a model law as a guideline. This is important in two ways: 1) It shows the integration of bankruptcy-related acts. Individual acts are integrated in order to restructure corporations more effectively. The international standardization of this law will facilitate more active international capital shift. Korea is in the process of establishing the Integrated Bankruptcy Law, considering the international trend. Also, a number of South Eastern countries are making efforts in this area. The resolution of NPLs expects to realize the establishment of the NPL market and related regulations. The resolution of NPLs has the trade-off problem as in other general regulations. The efficiency and speed in the process of the resolution of NPLs need to be ensured as well as the protection of rights for involved entities and the pursuit of legislative stability. Also, in the practical regulation, vested interest of the participants needs to be protected. Individual nations need to decide where they want to put more emphasis. The international standardization of relevant legislation in resolving NPLs could be pursued in various ways. Efforts for fundamental cooperation constitutes of the exchange of information and the introduction of different systems. Then, efforts need to be put in the establishment of similar regulations, transfer of model laws, establishment of international agreements. Mutual support program would be effective in the fund raising as well. Public organizations among different countries can exchange an MOU(Memorandum of Understanding) in the process of the international standardization of relevant legislation. An MOU have legal validity, similar to the cases of agreement. Generally, an MOU is made as a subsequent measure or a follow-up measures. However, an MOU among private organizations are not legally mandatory. The KAMCO in Korea have established a number of MOU with public organizations in many different countries. They constitute the following : 1) the resolution of NPLs through cooperative work; 2) Annual conference for the promotion of business exchanges; 3) Experience sharing for the resolution of NPLs; 4) cooperative work for establishing marketing strategies for the insolvent asset; and 5) the rule of confidentiality obligation, legal consequences, and termination of MOU. An international forum was held to pursue international cooperation among public NPL resolution agencies in Seoul, November, 2000. Representatives from 31 countries participated in the forum, including international organizations such as World Bank, Asia Development Bank, International Monetary Bank, and International Investment Bank. An international forum could be effective in the long-term development. With the cooperation and understanding among public NPL resolution agencies in different nations, it would be helpful to set up a system for exchanging asset information. International investors as well as domestic investors would like to collect information before entering the market. This is an approach taken by KAMCO in Korea, and it is a realistic way of pursuing the resolution.. The effectiveness of an MOU will be greater when Public organizations in different nations have realistic and practical rights within the government and financial institutions. The effectiveness of this method depends largely on whether the parties concerned on agreement have influential power in the international society. The Basel Committee on Banking Supervision, which is recognized as having international expertise, set up the BIS standard, and it exerts great influence in the world. In countries where public organizations play important roles, the MOU between public organizations would exert greater influence. In general, individual governments in the process of resolving NPLs establish public organizations on a temporary basis or increase the fund of existing organization. Therefore, governments need to play more active roles in resolving NPLs. Thus, cooperation among international financial institutions are required in order to facilitate mutual understanding and to establish more solid financial structures in turn. The systems to resolve NPLs usually take the forms of substantive laws or procedural laws. In substantive laws, the resolution of NPLs include the establishment of priority for loans, SPC(Special purpose companies), trust system, ABS, and tax support. These privilege clauses maintain the stability of existing legislation and effective resolution of NPLs at the same time. In procedural laws, it is possible to organize meetings of creditors or a general meeting of stockholders for the purpose of solving legal conflict. Also, it is possible to simplify the procedures of loan assignment. The movement of legislation in resolving NPLs usually take the form of special laws. Individual nations make legislation under their own circumstance. However, since the globalization of international economy develops rapidly, it is increasingly more important to consult foreign countries' legislation. Existing international organizations such as OECD, BIS, IMF or ADB could play important roles in the process. In addition, it is necessary to explore the possibilities of giving legal and mandatory power to the international regulations through multi-parties agreement. In establishing a model law, it would be necessary to examine different systems in many countries. This process entails the consideration of particular circumstances in which different nations operate. Then, it is desirable to prepare a number of model laws so that individual nations could select the best possible choice. Model laws could be developed according to the development stage of different NPLs markets. Model

Development of the High-Resolution Solar Radiation Model and Its Application on the Korean Peninsula

조일성 강릉원주대학교 일반대학원 2014 국내박사

GWNU (Gangneung-Wonju National University) one-layer solar radiation model is developed in order to resolve the lack of the vertical structure of atmospheric components and fast calculation with high horizontal spatial resolution. GWNU model is based on IQBAL and NREL methods and corrected by precise multi-layer LBL (Line-By-Line) model. Further, the amount of solar radiation reaching the surface by using 42 types of vertical atmospheric data as input data was compared with detailed models and one-layer models. One-layer solar radiation models were corrected depending on sensitivity of each input data (i.e. total precipitable water, ozone, mixed gas, and solar zenith angle). Global solar radiation was calculated by corrected GWNU solar model with satellites (MODIS, OMI and MTSAT-2), KLAPS model prediction data in Korea peninsula in 2010, and the results were compared to surface solar radiation observed by 22 KMA solar radiation sites. Calculated solar radiation annually accumulated showed highest solar radiation distribution in Andong, Daegu, and Jinju regions, meanwhile the observation data showed lower solar radiation in Daegu region compared to model result values. The GWNU (Gangneung-Wonju Nationa University) one-layer solar radiation model, which was corrected using line-by-line and vertical atmosphere data was developed based on the assumption that the ground surface is pane-parallel. GWNU_topo model was improved by taking into account the slope and aspect of the ground surface altitude, using the GWNU model where DEM (Digital Elevation Model) was applied. Using the GWNU_topo model applied the effects of the topography, the solar energy reaching the ground surface for the Korean Peninsula (resolution: 1 km × 1 km) and the Seoul region (resolution: 10 m × 10 m) was calculated. An analysis of the solar energy in relation to the characteristics of the ground surface was conducted. During the winter when the solar altitude is small, there was a big difference of the effect in relation to the topography due to the shielding of the solar energy (1 km × 1 km resolution : maximum ±10% or more, 10 m × 10 m resolution : maximum ±30% or more), while the difference during the summer was small. The higher the resolution for the topography data, the larger the deviation for solar energy. There was no difference in the maximum value of solar energy but due to the effects of the altitude and the slope, the monthly accumulated minimum value was 0.1 MJm-2.

Wavepath analysis and strategies for waveform inversion in the laplace domain

권정민 서울대학교 대학원 2017 국내박사

Laplace-domain waveform inversion (WI) is a technique for estimating long wavelength velocity models. The velocity model, estimated by Laplace-domain WI, is used as an initial velocity model for techniques such as frequency-domain and time-domain waveform inversion. These techniques are then used to develop high resolution velocity models used in subsurface imaging. Since frequency-domain and time-domain waveform inversion are sensitive to the initial velocity model, model resolution of Laplace-domain WI is an important factor in the overall velocity-estimation process. In addition, since the cost for obtaining the wavefield of the Laplace domain is large, it is necessary to improve the convergence rate and efficiency of Laplace-domain WI. Previous Laplace-domain WI studies have shown difficulty in analyzing model resolution and convergence rate due to insufficient understanding of the wavepath and its role in representing the relationship between the model parameters and seismic data. This study investigates the characteristics of the wavepath in the Laplace domain which have not been clarified in previous research. Through this study, we implement convergence rate, model resolution, and efficiency analysis for Laplace-domain inversion. By introducing the attenuation constant, which can be considered a Laplace constant in the spatial domain, we prove that the wavepath of the Laplace domain is a real exponential basis with the product of the attenuation constant vector and the position vector as an exponent. We also prove that the attenuation constant vector is a function of both the Laplace constant and the incident angle. From the numerical example, it can be confirmed that the attenuation constant depends on both the Laplace constant and the incident angle. In addition, this study shows that it is reasonable to apply the Gauss-Newton method to Laplace-domain WI for fast convergence. The wavepath of the Laplace domain is a real exponential function, which has a large condition number. The numerical example of the BP benchmark model demonstrates the effectiveness of the Gauss-Newton method in this Laplace-domain WI algorithm. We also prove that a wide range of incident angles is essential to obtain a high resolution model through Laplace-domain inversion. The relationship between the model resolution and the incident angle range explains why the model resolution decreases as the offset-depth ratio increases. Also, horizontal and vertical resolution changes, depending on the exploration environment, can be predicted. Finally, we propose an efficient Laplace constant selection strategy to improve the efficiency of Laplace domain inversion. The Laplace constants selected through the proposed method improve efficiency by maintaining continuity of the range of attenuation constants and by minimizing unnecessary repetition of attenuation constants. From the numerical example, it can be seen that the proposed Laplace constant selection strategy yields superior results in terms of both efficiency and accuracy, compared with the strategy of choosing Laplace constants at fixed intervals. This applies for both the simple-model and complex-model case, such as the SEG/EAGE salt dome model. 라플라스 영역 파형역산 알고리즘은 장파장 탄성파 속도 모델 추정 기술로서, 주파수 영역 및 시간 영역 파형역산과 같은 고해상도 속도 모델 추정 기술에 초기 속도 모델을 제공하는 용도로 사용된다. 주파수 영역 및 시간 영역 파형역산은 초기 속도 모델에 대단히 민감하기 때문에, 라플라스 영역 파형역산의 정확성은 전체 속도 모델 추정 과정에 있어서 대단히 중요한 요소이다. 또한 라플라스 파형역산에서 사용되는 라플라스 영역 파동장은 얻는 과정에서 많은 비용을 요구하기 때문에, 기술의 경제성 측면에서 라플라스 파형역산의 수렴속도 및 효율성은 역산 과정의 성패를 가르는 중요한 요소이다. 그러나 기존의 라플라스 파형역산에 대한 연구들에서는 모델 파라미터와 자료간의 관계를 나타내는 파동경로(wavepath)에 대한 고찰이 불충분한 관계로 모델 해상도 및 수렴속도 분석을 수행하는 데 어려움이 있었다. 본 연구는 기존의 연구에서 밝히지 못하였던 라플라스 영역의 파동경로의 성질을 규명하고, 이를 통해 기존의 연구에서 제대로 수행하지 못하였던 라플라스 영역 파형역산에 대한 수렴속도 및 모델 해상도, 그리고 효율성 분석을 수행한다. 먼저 공간 영역에 대한 라플라스 상수라 할 수 있는 감쇠 상수(attenuation constant)라는 개념을 도입함으로써 라플라스 영역의 파동경로가 근사적으로 감쇠 상수 벡터와 공간 벡터의 곱을 지수로 하는 실 지수함수 기저임을 증명한다. 이에 더하여 본 연구는 라플라스 영역의 파동 경로가 큰 조건수를 가지는 실 지수함수인 것을 통해, 빠른 수렴속도를 위해서는 라플라스 영역 파형역산에 가우스뉴턴법을 적용하는 것이 합리적임을 밝힌다. BP 벤치마크 모델의 수치 예제는 이러한 라플라스 영역 파형역산 알고리즘에서 가우스뉴턴법이 가지는 효용성을 증명해준다. 그리고 감쇠 상수 벡터가 라플라스 상수와 파의 입사 각도에 대한 함수임을 증명함으로써, 라플라스 영역 파형역산을 통해 높은 해상도의 모델을 얻기 위해서는 넓은 범위의 입사 각도가 필수적임을 밝힌다. 이러한 모델 해상도와 입사 각도 범위와의 관계는 오프셋-심도비(offset-depth ratio)가 증가함에 따라 모델 해상도가 낮아지는 이유를 설명해주며, 탐사환경에 따른 수평 및 수직 해상도의 변화 역시 예측할 수 있게 한다. 마지막 본 연구는 라플라스 영역 파형역산의 효율성을 향상시키는 방법으로 효율적인 라플라스 상수 선택 전략을 제안한다. 본 연구에서 제안하는 방법을 통해 선택된 라플라스 상수는 감쇠상수의 범위의 연속성을 유지시킴으로써 모델 해상도를 보장하며, 감쇠상수의 불필요한 중복을 최소화함으로써 효율성을 향상시킨다. 수치 예제로부터 제안된 라플라스 상수 선택 전략이 기존 연구에서 쓰여왔던 등간격으로 라플라스 상수 선택하는 전략에 비해 효율성 및 정확성의 두 가지 측면에서 월등한 결과를 산출해 내는 것을 확인할 수 있다.

Fast and Trustworthy Super-Resolution with VLM-Guided Diffusion Model : 비전-언어 모델 가이드를 통한 빠르고 신뢰할 수 있는 초고해상화 모델

성민규 경북대학교 대학원 2026 국내박사

Image Super-Resolution (SR) is a pivotal technique for reconstructing high-fidelity imagery from low-resolution inputs, underpinning critical applications ranging from urban planning and disaster response to scene text recognition. While diffusion models have emerged as the dominant paradigm for perceptual SR, their iterative nature imposes significant computational costs and high latency. Furthermore, existing pipelines often struggle to balance hallucination suppression with perceptual quality, frequently failing to preserve structural fidelity in remote sensing imagery or text readability in scene images. To address these challenges, this dissertation presents a unified framework that enhances the speed and trustworthiness of diffusion-based SR through modular guidance and structural decoupling. The contributions are threefold. First, from a systems perspective, H2-Cache introduces a hierarchical dual-stage caching mechanism with independent thresholds for structure and detail. By employing Pooled Feature Summarization (PFS), this approach significantly accelerates inference—achieving a 5.08× speedup with negligible perceptual degradation—by effectively capitalizing on the temporal redundancy of diffusion steps. Second, in the domain of remote sensing, a two-stage VLM-guided diffusion strategy is proposed. This method synthesizes a factually grounded foundation in the initial stage, followed by Vision-Language Model (VLM) textual prompts and ControlNet structural cues to enrich high-frequency details while maintaining geometric integrity. To rigorously evaluate this approach, the Segmentation Mask Similarity (SMS) metric is introduced, demonstrating balanced improvements in boundary preservation and perceptual quality across standard benchmarks. Finally, addressing the challenge of scene text readability, GLYPH-SR integrates OCR-specific constraints with generic semantic signals. By injecting these priors through a dedicated Text–SR fusion ControlNet (TS-ControlNet) and utilizing a four-regime synthetic training suite, the model successfully restores complex text features, such as logos and jersey lettering, yielding substantial gains in both OCR accuracy and SR metrics. Collectively, these components establish a coherent blueprint for fast and trustworthy super-resolution. By decoupling structure from detail and accelerating redundant computations, this work advances the state-of-the-art in semantic fidelity and operational efficiency, offering a robust direction for future research in generative image restoration.

Solving Real-World Visual Problems via Diffusion Models

장효영 광운대학교 일반대학원 2026 국내석사

In recent years, deep learning has led to significant performance gains in various computer vision tasks such as image segmentation and object detection. Their performance often heavily relies on large-scale, high-quality datasets. In practice, data scarcity caused by privacy and security constraints, as well as degraded image conditions such as low illumination and low resolution, lead to instability in training and poor generalization, limiting real-world applicability. This study proposes diffusion-based restoration and augmentation methodologies capable of achieving physically consistent and perceptually faithful image generation even under limited data conditions. Three key approaches are introduced. First, a Rician Gaussian diffusion model is proposed, designed to probabilistically simulate real pathological brain MRI by modeling the forward diffusion process. This allows the model to reflect abnormal noise distributions and structural deformations that commonly occur in clinical environments, thereby achieving higher realism and statistical consistency compared to Gaussian-only models. Moreover, it enables the precise synthesis of anomalous MRI data, improving model generalization and diagnostic reliability in data-scarce medical imaging domains. Second, a structure-aware low-light enhancement framework is presented, combining multi-objective losses including noise restoration, structural consistency, gradient, and total variation losses to achieve both illumination enhancement and preservation of structural and visual quality. Third, a patch-wise conditional diffusion model is proposed for efficient X-ray super-resolution. By applying Gaussian mask blending during image merging, the method preserves structural continuity across overlapping regions while enhancing fine anatomical details and computational efficiency. Experimental results demonstrate consistent improvements across various imaging domains. The proposed Rician Gaussian diffusion model achieved up to +2.4 dB PSNR and +3.7% SSIM gains over Gaussian-based diffusion models, with a 5–6% Dice score improvement in tumor segmentation. The structure- aware low-light enhancement framework achieved +0.54 dB PSNR and 13% lower LPIPS, effectively restoring illumination balance and sharpness under challenging low-light conditions. The X-ray super-resolution model improved PSNR by 1.8 dB and SSIM by 2.1%, maintaining structural coherence in fine details. Overall, this study demonstrates that diffusion models augmented with domain-aware priors, multi-objective optimization, and patch-based restoration strategies can effectively perform image restoration and augmentation in diverse domains, even under limited and noisy data conditions. 최근 영상 분할, 객체 탐지 등 다양한 컴퓨터 비전 태스크는 딥러닝 기술의 발전과 함께 비약적인 성능 향상을 이루었다. 이러한 모델의 성능은 대규모의 고품질 데이 터셋에 크게 의존하는 경향이 있다. 다만, 개인 정보 보호 및 보안상의 이유로 확보 가 어려운 희소 이미지 데이터나, 저조도 및 저해상도로 열화된 이미지 데이터는 학습의 불안정성과 일반화 능력의 한계를 초래하여 실제 환경에서의 활용성을 저해 한다. 본 연구는 제한된 데이터 환경에서도 물리적으로 일관되고 지각적으로 충실한 이미지 복원과 생성이 가능한 확산 기반 복원 및 증강 방법론들을 제안한다. 본 논 문은 세 가지 접근법을 제안한다. 첫째, 라이시안-가우시안 확산 모델을 제안하여 실제 이상 두뇌 자기 공명 영상을 확률적으로 모사하는 전방 확산 과정을 설계하였 다. 이를 통해 실제 임상 환경에서 발생하는 비정상적 노이즈 분포와 구조적 변형 을 반영할 수 있어, 단순한 가우시안 기반 모델 대비 더 높은 현실성과 통계적 일 관성을 확보하였다. 또한 제한된 데이터 환경에서도 이상데이터에 대한 정밀한 생 성을 통해 데이터 희소성이 심한 의료 영상 분야에서 모델의 일반화 성능과 진단 신뢰도를 향상시켰다. 둘째, 구조 인식 저조도 향상 프레임워크를 제안하여, 노이즈 복원, 구조적 일관성, 그래디언트, 전체 변동성 손실을 결합한 다중 목적 손실 함수 를 제안한다. 이를 통해 조도 향상과 함께 구조적 보존과 시각적 품질을 동시에 확 보한다. 셋째, 패치 단위 조건부 확산 모델을 제안하여 효율적인 엑스레이 초해상도 기법을 제안한다. 또한, 복원 이후 병합 이미지에 가우시안 마스크 블렌딩을 결합함 으로써, 중첩 영역 간의 구조적 연속성을 유지하면서 세밀한 해부학적 정보를 복원 하고 계산 효율성을 향상시켰다. 실험 결과, 제안된 프레임워크는 다양한 영상 도메인에서 일관된 성능 향상을 보였다. 라이시안-가우시안 확산 모델은 기존 가우시안 기반 확산모델 대비 PSNR이 최대 +2.4 데시벨, SSIM이 +3.7% 향상되었으며, 종양 분할에서 Dice 지표가 5–6% 개선되었다. 구조 인식 저조도 향상 모델은 PSNR +0.54 데시벨 상승과 LPIPS 13% 감소를 달성하여 저조도 환경에서의 조명 균형과 선명도를 효과적으로 복원하였다. 엑스레이 초해상도 복원에서는 PSNR이 1.8 데시벨, SSIM이 2.1% 향상되어 구조적 연속성을 성공적으로 유지하였다. 본 연구는 확산 모델을 활용하여 도메인 인식형 사전 지식, 다중 목적 최적화, 그리고 패치 기반 복원 최적화 프레임워크를 적용함으로써, 확산 모델이 다양한 도 메인에서 데이터가 제한된 환경에서도 영상 복원 및 증강 작업에 효과적으로 적용 될 수 있음을 입증하였다.

효율적인 의미 및 구조 분리를 통한 이중 경로 조건화 기반의 초해상화

김태환 세종대학교 일반대학원 2025 국내석사

최근 고화질 미디어 콘텐츠에 대한 수요가 급증하면서, 저해상도 이미지를 선명하게 복원하는 초해상화 기술의 중요성이 더욱 부각되고 있다. 그 러나 기존 CNN 및 트랜스포머 기반 모델은 고주파수 정보를 제대로 복원 하지 못해 흐릿한 화질을 생성하고, GAN 기반 모델은 학습의 불안정성으로 인해 다양한 형태의 왜곡이 발생하는 한계를 지닌다. 이를 보완하기 위해 등장한 확산 모델은 사실적인 질감 생성 능력에서 뛰어난 성능을 보였 지만, 저해상도 입력 정보를 단일 방식으로 처리하는 구조적 제약으로 인해 원본 구조가 왜곡되는 문제가 여전히 존재한다. 본 연구는 이러한 문제를 해결하기 위해, 저해상도 조건 정보를 의미 경로와 구조 경로라는 두 개의 독립적 흐름으로 분리하여 활용하는 새로운 이중 경로 조건화 구조를 제안한다. 제안하는 모델은 의미 경로를 통해 전역적 맥락을 심층 의미론적 단계에 적용하여 전체적인 문맥의 일관성을 보장하고, 구조 경로를 통해 지역적 세부 정보를 얕은 복원 단계에 직접 삽입하여 픽셀 수준의 정밀한 구조 복원을 유도하는 이원적 전략을 따른다. 제안 모델을 표준 벤치마크에서 평가한 결과, 트랜스포머 기반 모델에서 나타나는 흐릿한 화질 문제를 효과적으로 해소하면서 선명도를 유지하였고, 확산 모델에서 빈번히 발생하는 구조적 왜곡을 억제하였다. 특히 Mean Opinion Score(MOS) 기반의 주관적 화질 평가에서 제안 모델은 평가자들에게 비교 모델 대비 더 높은 점수를 기록하였다. 이는 제안 모델에서 저해상도 정보의 이원적 처리가 인간 지각 화질 향상에 실질적으로 기여함을 명확히 보여준다. 주요어 : Image Super-Resolution, Diffusion Model, Dual-path Conditioning, Perceptual Quality, Global Context, Local Detail With the surging demand for high-quality media content, Single Image Super-Resolution has become increasingly critical for restoring low-resolution images with high fidelity. However, traditional CNN and Transformer-based models often suffer from over-smoothed results in high-frequency regions due to their inability to fully recover fine details. Conversely, GAN-based models, while improving sharpness, frequently generate various forms of distortions due to training instability. To overcome these limitations, diffusion models have emerged as a powerful alternative with superior texture generation capabilities. Nevertheless, they face fundamental limitations arising from processing low-resolution (LR) information in a unified manner. This structural constraint often leads to "structural distortion" where the model probabilistically generates details that deviate from the original structure, even if the visual quality appears enhanced. To fundamentally address these challenges, this study proposes a novel Dual-Path Conditioning architecture that disentangles LR condition information into two independent streams: global context and local details. The proposed model adopts a dual strategy where the global context is injected into the deep semantic stages of the network to guide overall consistency, while local details are applied directly to the shallow restoration stages to facilitate precise pixel-level structural recovery. Extensive experiments on standard benchmarks demonstrate that the proposed model effectively resolves the blurriness issues of Transformer-based models while maintaining sharpness, and simultaneously suppresses the structural distortions frequently observed in diffusion models. Notably, in subjective quality evaluations based on Mean Opinion Score (MOS), evaluators rated the proposed model as having the most natural quality, free from both distortions and blurriness, demonstrating a distinct advantage over existing methods. This MOS-centric analysis clearly indicates that the dual-path processing of LR information significantly contributes to the enhancement of human perceptual quality. Keywords : Image Super-Resolution, Diffusion Models, Dual-Path Conditioning, Perceptual Quality, Global Context, Local Detail

Estimation of High-Spatial Resolution of Ground-Level Ozone, Nitrogen Dioxide, and Carbon Monoxide in South Korea During 2002-2020 Using Machine-Learning Based Ensemble Model

권도훈 서울대학교 대학원 2023 국내석사

Backrgound : Long-term exposure to ozone (O3), nitrogen dioxide (NO2), and carbon monoxide (CO) is known to cause various diseases and increase mortality. For that reason, estimating ground-level O3, NO2, and CO concentrations with a high spatial resolution is crucial for assessing the health effects associated with these air pollutants. However, related studies are limited in South Korea. This study aimed to develop machine learning-based models to predict the monthly O3 (average of daily 8-hour maximums), NO2, and CO at a spatial resolution of 1 km × 1 km across South Korea from 2002 to 2020. Methods : Approximately 80% of the monitoring stations were used to train the three machine learning models (random forest, light gradient boosting, and neural network) with a 10-fold cross-validation, and 20% of the monitoring stations were used to test the model performance. The author also applied ensemble models to integrate the variation in predictions among the models. Multiple predictors with satellite-based remote sensing data, inverse distance weighted ground-level air pollutants, land use variables, reanalysis datasets for meteorological variables, and regional socioeconmoic variables collected from various databases were included in the prediction model. Results : For O3, the overall R2 of the ensemble model was 0.841 during the entire study period. Urban areas showed a better model performance (R2 = 0.845) than rural areas (R2 = 0.762). For NO2, the highest overall R2 was 0.756, which best fit in autumn (R2 = 0.768). For CO, the overall R2 value was 0.506. This study provides high spatial resolution monthly average O3 and NO2 estimates with excellent performance (R2 > 0.75). Conclusion : The author’s predictions can be used to analyze the spatial patterns in pollutants in relation to population characteristics and studies on the health effects of long-term exposure to air pollution using geocode-based health information and local health data. 연구배경 : 오존(O3), 이산화질소(NO2), 일산화탄소(CO)에 장기간 노출되면 각종 질병을 유발하고 사망률을 높이는 것으로 알려져 있다. 그렇기에, 고해상도로 지표면 O3, NO2, CO 농도를 추정하는 것은 이러한 대기오염물질과 관련된 건강 영향을 평가하는 데 매우 중요하다. 하지만, 장기간에 걸쳐 고해상도로 가스상 대기오염물질(O3, NO2, CO)를 추정한 연구는 국내에서 아직 진행된 바가 없다. 따라서, 본 연구는 2002년부터 2020년까지 대한민국 전역에서 1km × 1km의 공간해상도로 월별 O3(일평균 8시간 최대치), NO2, CO를 머신러닝 기반 모델 및 그들의 앙상블 모형을 통해 예측하고자 한다. 연구방법 : 3가지 머신러닝 모델(랜덤 포레스트, 라이트 그래디언트 부스팅, 신경망)의 최적의 파라미터를 찾기 위해 모니터링 스테이션의 약 80%를 훈련 데이터로 사용하였고, 10-fold 교차검증을 통해 훈련 데이터 내에서 훈련/평가 단계를 거쳤으며, 나머지 모니터링 스테이션의 20%를 모델 평가에 사용하였다. 여기에 추가로 머신러닝 모델 간의 예측 변동을 통합하기 위해 앙상블 모델을 적용했다. 데이터에는 위성 기반 원격 감지 데이터, 역거리 가중치 기반 대기오염농도, 토지 이용 변수, 기상 재분석 자료, 다양한 데이터베이스에서 수집된 지역 사회경제적 변수 등이 포함되었다. 연구결과 : O3의 경우, 전체 연구 기간 동안 앙상블 모델의 R2가 0.841을 기록했으며, 도시 지역이 농촌 지역(R2 = 0.762)보다 우수한 예측 성능(R2 = 0.845)을 보였다. NO2의 경우, 앙상블(평균) 모델의 R2가 0.756으로 가장 높았으며, 계절로 보면 가을에 예측 성능이 가장 높았다(R2 = 0.768). CO의 경우, R2가 0.506 을 기록했다. 본 연구는 O3 및 NO2 에서 R2 > 0.75 으로 높은 예측력의 고해상도 월평균 추정치를 제공한다. 결론 : 본 연구에서 얻어진 대기오염 추정 결과는 인구 특성과 관련된 가스상 대기오염물질의 공간 패턴을 분석하거나, 위치 기반 건강 정보와 행정구역 단위 건강 데이터와 엮여서 장기간 대기오염 노출의 건강 영향을 평가하는 연구에 사용될 수 있을 것으로 기대된다.

Development of high-resolution regional model for simulating CO2 mole fraction over East Asia and the effect of atmospheric field, VPRM parameter, and anthropogenic emission inventory

서민경 Graduate School, Yonsei University 2021 국내석사

본 연구에서는 고해상도의 지역 단위 이산화탄소를 모의하기 위해 Weather Research and Forecasting – Chemistry (WRF-Chem) 모델을 구축하였다. WRF-Chem 모델 내에서 이산화탄소는 식생, 해양, 인위적 배출, 배경 농도로 세분화되며, 해양, 인위적 배출 이산화탄소 농도는 각각의 인벤토리로부터 제공된다. 식생 이산화탄소 농도는 WRF-Chem 모델 내에 내장된 Vegetation Photosynthesis and Respiration Model (VPRM)을 통해 계산된다. 이산화탄소는 대기 중의 다른 화학 종과 달리, 바람에 의한 수송의 영향을 많이 받는다. 따라서 정확한 이산화탄소 농도가 모의되기 위해서는 바람장이 정확히 모의되어야 한다. 또한, 식생 이산화탄소를 계산하기 위한 VPRM에서는 Gross Ecosystem Exchange (GEE) 값과 Respiration (R) 값을 계산하기 위해 특정 매개변수 값이 필요하며, 해당 매개변수 값은 실험 지역에 맞게 최적화되어야 한다. 따라서, 본 연구에서는 동아시아 지역 고해상도의 이산화탄소를 모의함에 있어 최적의 기상장 및 VPRM 매개변수를 선정하는 실험을 수행하였다. 또한, 결정된 최적의 실험 설계를 바탕으로 2018년 1년 기간에 대해 동아시아 지역 이산화탄소 농도를 모의하여, 그 결과를 분석하였다. 기상장 및 VPRM 매개변수 선정을 위해, 기상장으로는 National Center for Environmental Prediction (NCEP)의 Final analyses (FNL)과 fifth generation European Centre for Medium-range Weather Forecasts (ECMWF) atmospheric ReAnalysis (ERA5)를, VPRM 매개변수로는 미국 지역을 대상으로 최적화된 US table과 중국 지역을 대상으로 최적화된 Li table (Li et al., 2020) 및 Dayalu table (Dayalu et al., 2018)을 고려하였다. 실험은 2018년 3월 한 달 기간에 대해 2가지 기상장과 3가지 VPRM table을 조합한 6개 실험이 수행되었다. 기상 및 이산화탄소 관측을 이용하여 검증한 결과, FNL 기상장과 US table의 조합이 동아시아 이산화탄소 농도 모의에 가장 적합하였다. 지상 이산화탄소 관측과의 검증 시, TAP (태안), AMY (안면도), KIS (기사이) 관측지에서 이산화탄소 농도가 매우 과대 모의되었는데, 세 관측지 모두 강한 인위적 이산화탄소 배출량이 존재하는 지역 인근에 위치하고 있다. 따라서, 기존 실험에서 사용되었던 Emission Database for Global Atmospheric Research (EDGAR) 인벤토리에서 Open-source Data Inventory for Anthropogenic CO2 (ODIAC) 인벤토리로 변경하여 2018년 3월 한 달 기간 동안 실험을 수행하였다. 분석 결과, ODIAC으로 변경한 실험 결과에서 TAP, AMY, KIS 지역의 이산화탄소 농도 과대 모의 정도가 크게 감소하여, ODIAC 인벤토리가 동아시아 이산화탄소 모의에 보다 적합하였다. 최적화된 WRF-Chem 실험 설계로 2018년 1년 기간에 대해 동아시아 지역 고해상도 이산화탄소 농도 모의를 수행 후 계절 별로 분석을 수행하였다. 분석 결과, 이산화탄소 농도가 관측과 비교하여 적절히 모의됨을 확인할 수 있었다. 또한 겨울, 봄, 가을철에는 이산화탄소 농도가 낮은 지역 (중국 북부 지역)에서 이산화탄소가 과대 모의되었으며, 이산화탄소가 높은 지역 (중국 내륙 지역)에서 이산화탄소가 과소 모의되었다. 본 연구에서는 동아시아 지역 고해상도 이산화탄소 모의를 위한 WRF-Chem 모델을 구축하고 최적의 실험 설계를 제시하였으며, 2018년 1년 기간에 대해 신뢰성 있는 고해상도 이산화탄소 농도가 모의됨을 확인하였다. 본 연구를 통해 이산화탄소 농도 추정의 불확실성이 큰 동아시아 지역의 탄소 순환 이해도를 증가시킬 수 있을 것으로 기대된다. In this study, the Weather Research and Forecasting – Chemistry (WRF-Chem) model, a high-resolution regional model, was developed to simulate CO2 concentrations in East Asia. In the WRF-Chem, CO2 concentrations are subdivided into 4 components: anthropogenic, oceanic, biogenic, and background. Anthropogenic and oceanic CO2 concentrations were provided from each emission inventory data. Biogenic CO2 concentration was calculated from Vegetation Photosynthesis and Respiration Model (VPRM) combined with WRF-Chem. CO2 is more affected by wind than other chemical components in the atmosphere. Thus, the wind speed and direction need to be simulated accurately to have accurate CO2 concentrations in the model. In addition, when calculating the gross ecosystem exchange (GEE) and Respiration (R) in the VPRM, some parameter values need to be provided and optimized for the experimental area. Therefore, in this study, several experiments were conducted to determine the optimal atmospheric field and VPRM parameter values which are appropriate for East Asia. With the optimal atmospheric fields and VPRM parameters, the CO2 concentrations in East Asia for a year of 2018 were simulated and compared with the observed CO2 concentrations. For the atmospheric field, the National Center for Environmental Prediction (NCEP) Final analysis (FNL) and the European Center for Medium-range Weather Forecasts (ECMWF) 5th ECMWF ReAnalysis (ERA5) were used as the initial and boundary conditions of the WRF-Chem. As the VPRM parameter tables, the US region table provided in the WRF-Chem and the tables specified in Li et al. (2020) and Dayalu et al. (2018) for parts of Asia were used. The experimental period was the one-month period of March 2018, and 6 experiments were performed combining 2 meteorological fields and 3 VPRM parameters. Based on the comparison with meteorological and CO2 observations, the combination of FNL data and US table is appropriate for the simulation of CO2 concentrations in East Asia. When compared with the surface CO2 observations, the modeled CO2 concentrations in Tae-ahn Peninsula (TAP), Anmyeon-do (AMY), and Kisai (KIS) are highly overestimated. These three observation sites are located near areas of strong anthropogenic CO2 emissions. To reduce the overestimation in CO2 simulations in above three sites, the emission inventory was changed from the Emission Database for Global Atmospheric Research (EDGAR) inventory to Open-source Data Inventory for Anthropogenic CO2 (ODIAC) inventory for the one-month period (i.e., March 2018) simulations of CO2 concentrations. By using the ODIAC inventory, the simulated CO2 concentrations become more similar to the observed CO2 concentrations in most verification sites compared to those with EDGAR inventory, which implies that the ODIAC inventory is appropriate for CO2 simulations in East Asia. With optimal experimental settings of FNL data, US table, and ODIAC inventory, one-year simulations in 2018 were performed to investigate the variability and performance of high-resolution CO2 concentration simulations in East Asia. Overall, the monthly averaged CO2 concentrations are appropriately simulated compared with observations. In winter, spring, and autumn, the modeled CO2 concentrations are underestimated in areas with high observed CO2 concentrations and are overestimated in areas with low observed CO2 concentrations. In this study, the WRF-Chem with an optimal experimental setup was developed for high-resolution simulations of regional CO2 concentrations in East Asia. With the developed model framework, the reliable high-resolution CO2 concentrations were simulated in East Asia for a year of 2018. The results of this study are expected to increase the understanding of the carbon cycle in East Asia where the estimation of CO2 concentrations is highly uncertain.