저시력인을 위한 화면확대 프로그램의 효율적 인터페이스 설계를 위한 제언

이수정 조선대학교 2004 국내석사

딥러닝 기반 전력 계량데이터 결측 보정 모델에 대한 연구

권혁록 조선대학교 대학원 2022 국내박사

Due to global warming, abnormal climates such as heavy snow, heatwaves, forest fires, and typhoons are occurring in many places around the world. Recognizing the seriousness of the climate change problem, the international panel on climate change (IPCC) general meeting in Songdo, Incheon in 2018 suggested that carbon dioxide emissions should be reduced by at least 45% compared to 2010 and achieved by 2050. Korea also prepared the "2050 Carbon Neutral Promotion Strategy" in December 2020. Five basic carbon-neutral directions: ①Expanding the use of clean electricity and hydrogen ②Improving innovative energy efficiency in connection with digital technology ③Promoting the development and commercialization of future carbon-free technologies ④Sustainable industrial innovation with a circular economy Promotion ⑤strengthening the carbon absorption function of nature and ecology such as forests, tidal flats, and wetlands were suggested. After all, the most basic thing to reduce carbon emissions is to reduce energy consumption. In order to save energy, it is possible to quickly build AMI and to reduce the active energy of consumers by supporting advanced time-based rate plans such as Time Of Use (TOU), Critical Peak Pricing (CPP), and Real Time Pricing (RTP) which are various additional services through two-way communication between AMI infrastructure. KEPCO has also tried to build an AMI infrastructure across 22.5 million units by 2020, but about 10 million units have been built and operated, and the construction will be completed in the next few years. In addition, the government announced in 2020 that it would distribute AMI to 5 million apartments in its detailed task, “Building an Intelligent Smart Grid for Energy Management Efficiency”, through the announcement of the “Korean Version of the New Deal”. As the spread of AMI expands, various new services are emerging, and through this, they are taking a step further toward the goal of carbon neutrality, a government policy. In order to provide additional services using power metering data, it is essential to acquire measurement data well from the electricity meter. However, due to the limitations and various factors of the communication network that makes up the AMI, there are about 2-5% missing values. In order to improve the quality of AMI additional services, it is necessary to predict and provide missing data. Therefore, various algorithms are being studied and applied to predict missing values of time series data generated by smart meters. However, if the accuracy is not high and the missing section is prolonged, the error increases and quality service cannot be provided. The power usage data is not a general time series data prediction, but has a continuously increasing cumulative power usage value, so the cumulative power usage value should be predicted and corrected. Among the methods frequently used in the time series field so far, this paper identified their strengths and weaknesses through previous studies such as linear correction, similarity-based correction, Autoregressive Integrated Moving Average(ARIMA) prediction correction, and LSTM(Long Short-Term Memory) prediction correction. However, it was not appropriate to use the previously studied correction methods to predict cumulative power usage data. In particular, while simply predicting and correcting the data, there was an error that some data decreased the cumulative power usage value that appeared after correction. To solve these problems, this paper proposes a model that applies weights to a deep learning-based CNN-LSTM combination model as a hybrid method that combines the advantages of linear correction method and correction method using LSTM deep learning algorithms superior to general methods used in time series. To conduct the study, the study was conducted in the order of Business Understanding, Data Understanding, Data Preparation, Model Development, and Evaluation according to the standards of CRISP-DM(Cross Industry Standard Process for Data Mining) methodology. Through previous studies, the AMI infrastructure was first understood, the cause and ratio of data missing were identified, the characteristics and patterns of power usage data were identified through data analysis, and data was prepared through preprocessing. In order to improve the previously studied long short-term memory deep learning algorithm (LSTM), a deep learning model was created that combines a convolutional neural network (CNN) and a short-term memory circulating neural network (LSTM). The results predicted by the CNN-LSTM combined deep learning model were predicted by applying weights, which are the ratio of each section to the total amount of missing sections. And finally, the final cumulative power usage data was predicted by summing the cumulative power usage value before missing and the hourly power usage value. When comparing and analyzing the experimental results of the preceding correction method and the proposed correction method, the correction method proposed in this paper was ‘0.218447’, which was about 40 times better than the ARIMA predictive correction method, and the square root mean square error (RMSE) was 6 times better. Even when compared with Mean Absolute Percentage Error(MAPE) value, the correction method proposed in this paper was ‘0.009899’, which was the best at about 45 times more than the ARIMA prediction correction method ‘0.448682’. In addition, the correction method proposed in this paper did not reduce the cumulative power usage value because the weight was applied to the total amount of missing sections, and even if there were many errors in predicting power usage over time, stability was very high. When analyzing the experimental results according to the length of missing, the linear correction method was generally simpler and better than various methods with good performance in the time series field. The linear correction method had the highest accuracy until the number of data in the missing section was less than 7. Therefore, when applying to the AMI system, it is effective to determine the missing length first and use the linear correction method if the missing length is less than 7, and in the above case, it would be better to use the CNN-LSTM combination weighting correction method proposed in this paper. In the future, if the proposed model was predicted and corrected with deep learning algorithm simply with time series data of power usage, research to increase its accuracy using more input values is needed to learn deep learning. In particular, since power usage has a very high correlation with climate, it is necessary to further increase accuracy by adding weather information. 지구 온난화로 폭설, 폭염, 산불, 태풍 등 이상기후 현상이 세계 여러 곳에서 나타나고 있다. 이에 국제사회는 기후변화 문제의 심각성을 인식하고 이를 해결하기 위해 2015년 파리기후변화협정, 2018년 인천 송도에서 개최된 IPCC(Intergovernmental Panel on Climate Change) 총회에서 지구온난화 1.5℃ 목표의 과학적 근거마련 등 2030년까지 이산화탄소 배출량을 2010년 대비 최소 45% 이상 감축하여야 하고, 2050년경에는 탄소중립(Netzero)을 달성하여야 한다는 경로를 제시했다. 우리나라도 2020년 12월 '2050 탄소중립 추진전략'을 마련했다. 여기에 탄소중립 5대 기본방향으로 ①깨끗하게 생산된 전기․수소의 활용 확대 ②디지털 기술과 연계한 혁신적인 에너지 효율 향상 ③탈 탄소 미래기술 개발 및 상용화 촉진 ④순환경제로 지속가능한 산업 혁신 촉진 ⑤산림, 갯벌, 습지 등 자연․생태의 탄소 흡수 기능 강화 등이 제시됐다. 결국 탄소배출량을 줄이기 위해서 가장 기본이 되는 것이 에너지 사용량을 줄이는 것이다. 에너지 절약을 위해서 AMI구축을 서두르고 AMI 인프라를 통해 소비자와 전력회사 간 양방향통신으로 다양한 부가서비스인 계시별 요금제(TOU; Time Of Use), 수요관리형 선택요금(CPP; Critical Peak Pricing), 실시간 요금제(RTP; Real Time Pricing) 등 고도화된 Time-based 요금제 지원을 통해 수용가의 능동적인 에너지 절감 참여 유도가 가능하다. 한국전력도 2020년까지 2,250만호 전체에 AMI 인프라를 구축하기 위해 노력했지만 약 1,000만호를 구축 운영 상태이며, 앞으로 수년 내에 구축이 완료될 것이다. 또한 정부는 2020년 “한국판 뉴딜” 발표를 통해 그 세부과제인 “에너지관리 효율화 지능형 스마트그리드 구축”에서 아파트 500만호를 대상으로 AMI를 보급하겠다고 발표했다. AMI 보급이 확대되면서 여러 가지 새로운 서비스들이 생겨나고 있고, 이를 통해 정부정책인 탄소중립 목표에 한걸음 더 나아가고 있다. 전력 계량데이터를 활용한 부가 서비스들을 제공하기 위해선 필수적으로 전력량계로부터 계량데이터들을 잘 취득하여야 한다. 그러나 AMI를 구성하고 있는 통신망의 한계 및 여러 가지 요인으로 인해서 계량데이터의 결측이 2~5% 가량 발생하고 있다. AMI 부가 서비스의 품질을 높이기 위해서는 결측 데이터를 예측하여 제공하는 것이 필요하다. 따라서 스마트미터에서 생성한 시계열 데이터의 값을 예측하기 위해 여러 가지 알고리즘들을 연구하고 적용하고 있다. 하지만 그 정확도가 높지 않고 결측 구간이 길어지면 오차가 커져 양질의 서비스를 제공할 수가 없다. 전력 사용량은 일반적인 시계열 데이터 예측이 아니라 계속 증가하는 누적 전력사용량 값을 가지고 있고, 따라서 누적 전력사용량 값을 예측해서 보정해야 한다. 지금까지 시계열 분야에서 자주 이용되는 방법들 중에 본 논문에서는 선형보정법, 유사도 기반 보정법, ARIMA(AutoRegressive Integrated Moving Average) 예측 보정법, 장단기메모리순환신경망(LSTM; Long Short-Term Memory) 예측 보정법 등을 선행 연구를 통해 그 장·단점을 파악하였다. 하지만 선행 연구된 보정법들을 누적 전력사용량 예측에 사용하는 것은 적합하지 않았다. 특히 단순히 데이터를 예측 후 보정하면서 일부 데이터들이 결측 후 나타나는 누적 전력사용량 값이 감소되는 오류가 발생했다. 본 논문에서는 이러한 문제들을 해결하기 위해 시계열에 사용되는 일반적인 방법들보다 선형보정법의 장점과 시계열 분야에서 우수한 LSTM 딥러닝 알고리즘을 이용한 보정법의 장점을 결합한 하이브리드 방법으로 딥러닝 기반 CNN-LSTM결합 모형에 가중치를 적용한 모델을 제안하였다. 연구를 수행하기 위해 CRISP-DM(Cross Industry Standard Process for Data Mining) 방법론의 표준에 따라서 업무 이해(Business Understanding), 데이터 이해(Data Understanding), 데이터 준비(Data Preparation), 모델 개발(Modeling), 평가(Evaluation) 순으로 연구를 진행하였다. 선행 연구를 통해서 AMI 인프라를 먼저 이해하고, 데이터 결측이 발생하는 원인, 비율 등을 파악했으며, 데이터 분석을 통해서 전력사용량 데이터들의 특징 및 패턴을 파악하고, 전처리를 통해서 데이터를 준비하였다. 선행 연구된 장단기메모리순환신경망(LSTM; Long Short-Term Memory) 딥러닝 알고리즘을 개선하고자 합성곱신경망(CNN; Convolutional Neural Network)과 장단기메모리순환신경망(LSTM)를 결합한 딥러닝 모형을 만들었고, CNN-LSTM결합 딥러닝 모형의 예측 결과에 결측 구간의 총량 대비 구간별 비율인 가중치를 적용하여 시간별 구간 전력사용량을 예측하였고, 마지막으로 결측 이전 누적 전력사용량 값에 시간별 구간 전력사용량을 합산하여 최종 누적 전력사용량을 예측하였다. 선행 보정법과 제안된 보정법의 실험 결과를 비교 분석했을 때 평균제곱오차 (MSE)는 본 논문에서 제안한 보정법이 ‘0.218447’로 ARIMA 예측 보정법보다 약 40배 이상 좋았고, 제곱근 평균제곱오차(RMSE)도 6배 이상 좋았다. 평균절대비오차(MAPE) 값으로 비교 시에도 본 논문에서 제안한 보정법이 ‘0.009899’ 으로 ARIMA 예측 보정법 ‘0.448682’ 보다 약 45배 이상으로 정확도가 가장 우수했다. 또한 본 논문에서 제안한 보정법은 결측 구간의 총량 대비 구간별 비율인 가중치를 적용하였기 때문에 누적 전력사용량 값이 감소되는 현상도 발생하지 않았고, 시간별 구간 전력사용량 예측 오차가 많이 발생하더라도 총 사용량 내에서 예측을 하게 되므로 안정성이 매우 높았다. 결측 길이에 따른 실험결과를 분석했을 때 일반적으로 시계열 분야에서 성능이 좋은 여러 가지 방법들보다 선형보정법이 간단하면서도 성능이 좋았다. 결측 구간의 데이터 개수가 7개 미만 일 때 까지는 선형보정법이 정확도가 가장 높았다. 그래서 AMI시스템에 적용 시에는 결측 길이를 먼저 판단하여 결측 길이가 7개 미만일 경우는 선형보정법을 사용하는 것이 효과적이고, 이상일 경우에는 본 논문에서 제안한 CNN-LSTM결합 가중치적용 보정법을 사용하는 것이 좋을 것이다. 향후, 제안된 모델이 단순히 전력사용량의 시계열 데이터만을 가지고 딥러닝 알고리즘으로 예측하여 보정하였다면, 딥러닝을 학습시키기 위해서 더 많은 입력값을 사용하여 그 정확도를 높이는 연구가 필요하다. 특히 전력사용량은 기후와 그 연관성이 매우 높기 때문에 날씨정보를 추가하여 정확도를 더 높일 필요가 있다.

단어 간 의미적 연관성을 고려한 개선된 문서 요약 방법 연구

차준석 조선대학교 산업기술융합대학원 2016 국내석사

Along with the recent development and distribution of smart devices, data contained in documents appearing on the internet are increasing exponentially. As documents are increasing exponentially, only the titles and main points are shown to users as a solution to figure out information in a document they want. Only with brief contents, however, users may find it difficult to get the information of a document they want. If document wanted by a user is expressed in an accurate form, it will be helpful when trying to find out necessary information. Document summarization is referred to eliminating redundancies and producing condensed information while maintaining consistency in collected documents. Automatic document summarization technology is to process a large amount of documents automatically and efficiently by extracting main sentences from a document using a computer and eliminating overlapped contents. To summarize a document efficiently, the present study uses a text rank algorithm. The text rank algorithm expresses sentences or keywords as a graph and employs the peaks or main lines of a graph to figure out semantic correlation between words and sentences and understand the importance of sentences. Through this, it goes through a process to extract the superordinate keywords of sentences and extract main sentences based on those keywords. To go through a process to extract main sentence s, word grouping is done. For word grouping, a particular weighting scale is used to screen sentences having a high weighted value. Based on the screen sentences, main sentences are extracted, and the document is summarized. This study has proven that it shows higher performance than any of the document summarization methods previously examined and performs summarization more efficiently.

단어 의미적 연관성을 고려한 개선된 어휘체인기반의 자동 문서요약 방법

택미얏린 조선대학교 2016 국내석사

문서를 요약한다는 것은 그 문서의 일관성을 유지하면서 중복을 제거하고, 응축된 정보를 생산하는 것을 말하며, 자동문서 요약 기술은 컴퓨터를 사용해서 문서 내 중요한 부분을 유지하고, 중복된 내용을 제거함으로써 처리하고자 하는 대용량의 문서를 자동적이고 효율적으로 처리하는 방법을 말한다. “어휘 결합(Lexical Cohesion)”은 어휘의 관계(상하 어휘관계, 유의 어휘관계 등)를 바탕으로 하나의 문서 내의 등장하는 단어와 단어 사이의 관계를 분석하는 방법이다. 이러한 “어휘 결합 관계” 는 어휘 사슬(Lexical Chain)을 이용하여 나타낼 수 있다. 어휘 사슬은 자연언어처리(Natural Language Processing) 및 정보검색(Information Retrieval)기술에 다양하게 활용되고 있으며, 적절한 후보키워드를 추출하는 것이 프로그램의 성능을 좌우하기 때문에, 어휘 사슬을 구성하기 위해 적절한 후보키워드를 추출하는 것이 가장 중요한 작업이라고 할 수 있다. 본 연구는 단어의 의미적 연관성을 고려하여 구성한 후보키워드의 어휘사슬을 기반으로 개선된 자동문서 요약방법에 관한 연구로, 효율적인 어휘 사슬을 구성하기 위해 새로운 키워드추출방법을 제안하였다. 본 논문에서 문서 내의 키워드 추출을 위해 “제목에 등장하는 단어”, “문서의 첫 번째 문장에 등장하는 단어”, “TF-IDF 가중치가 높게 측정되는 단어” 세 가지의 키워드 특징을 정의하였으며, 키워드가 갖는 조건부 확률 값을 활용해 전이 행렬(transition matrices)을 생성함으로써, 마르코프 연쇄(Markov Chain)에 적용을 통해 후보키워드를 추출한다. 추출된 후보키워드는 워드넷(WordNet) 상에서 정의된 단어의 상하위어 관계, 동의어 관계를 고려하여 후보 키워드 간의 연결을 통해 어휘 사슬을 구성하였으며 이를 통해 자동문서 요약을 수행하게 된다. 본 논문의 실험결과에 따르면, 제안한 방법에 의해 추출한 후보키워드로 어휘 사슬을 구성하는 것이 문서 내 모든 명사구에 대한 어휘 사슬을 구성하는 것보다 향상된 성능을 보였으며, 더욱 효율적으로 문서를 요약할 수 있음을 증명하였다. Summarization is a challenging task that need to understand the content of the document to determine the importance information of the text. Automatic Summarization is the procedure of lessening a text document using an intelligent system with complex algorithms to form a short summary of it which retains the most important information of the text. Lexical cohesion is a way identifying connected portions of the text according to the relations between the words in the document or text. Lexical cohesive relations between words in a document can be described using lexical chains. Lexical chains are applied in various Natural Language Processing (NLP) and Information Retrieval (IR) applications. The fundamental task to perform in constructing lexical chain is to extract the best appropriate candidate keywords of the text as the chains which contain the salient portions of the document are relied on them. Thus, we extend our research on automatic keyword extraction for extracting candidate terms of the given text before approaching to summarization. Keywords are a set of keywords or keyphrases that capture the primary information or topic discussed in the text. Keywords are widely used to define queries in IR systems as they are easy to define, revise, remember, and share. Furthermore, keywords are essential Search Engine Optimization (SEO) elements for every search engines, yet they are matched against with users’ search query keywords. They can empower document browsing by providing a short summary, improve information retrieval, and be employed in generating indexes for a large text corpus. In current thesis, we proposed a new approach for automatic text summarization using lexical chain with semantic relatedness keywords. In contrast, the new method of extracting keywords is also implemented for constructing the efficient lexical chain. Instead of constructing a lexical chain with every noun phrases in the text, the result of our experimental results show that the extracted candidate keywords of the text delivers a better efficient summary with a better performance. Thus, we have built a system to extract the promising keywords from the text which is based on the characteristics of the manually assigned keywords. The system consists of a generator to produce the possibility distributions of three distinct features of assigned keywords; the occurrence of a term in title, the occurrence of a term in a first sentence, and the higher score of average TF.IDF score of a term. Then, those distributions are applied to Markov chain process of three stages to assign the label for each N-gram term in the text. Extracted unigram candidate keywords are selected to build the lexical chains of the text. There are also three distinct relation criteria to connect the terms according to their relationships to other candidate keywords using WordNet; hypernym, hyponym and synonym. Then we apply the method to score and extract the salient portions of the document. Our experimental results prove that the efficient summaries can be extracted for the above tasks.

Link grammar를 이용한 도메인 온톨로지 확장 방안

윤병수 조선대학교 2010 국내석사

Ontology is constructed with concept, definition, and these relation. lately, much of data, ontology do not support reasoning information to users. so, there are so many necessary for ontology population. therefore many studies in ontology population. but most of study, they used manually extraction to concept, relation, properties. this method spend a lot of time and money to data mining. so, to solve this problem, automatic ontology population have studied. it save much of time, money. however it requires another knowledge-base or thesaurus. and they dependent on its source(thesaurus, knowledge-base, etc) In my study, I deal with ontology population which are not using another famous dictionary, but using Link grammar and infobox of wikipedia. Link grammar is a syntactic parsing theory of English. I analyze link pattern to determine what link pattern will be candidate of relation-concept and extracted triples. then added weight value to each triples. in the result, i got relation and concepts, in order to weight value. weight values are purpose to extract good triples. then i apply infobox, navbox to classify relation, concepts. First, I gather biology documents in wikipedia. then to get a body part of wiki-document, apply stampling process. and extract terminologies, which make database to extract important sentences. i set the process for terminology extraction(tagging, tokenize, extraction). and verify terminologies. then i select important sentences to apply Link grammar. I define 7-patterns to get concept-relation triple. after i find good triple through the PMI value and TF value. At last, i classify concept with infobox, navbox and classify relation with pre-defined classify table which defined Relation hierarchy. i make metadata with this properties, then visualization it. There are some error in my study, it occurs in wrong Pos tagging, stopword interruption. anaphora relosolution. but visualization works well, and it's possible to extract more than 1 relations in 1 sentence. finally, I will study about named entity recognization to finding correct triples .

지능적 문서 분석을 위한 개선된 WSD 방법 연구

최동진 조선대학교 2015 국내박사

Natural language is a communication system that was created by the human evolutionary process, in order to write and share information with others. Starting from the simple methods (gestures or facial shapes), natural language has been evolved into highly scientific and intelligent language systems. With the advent of a paper, people started to produce many kinds of texts however, the internet and smart devices has changed entire world. Due to the simplicity and mobility of smart devices, documents on the web have been dramatically increased during past few years. The problem is that handling huge amount of web documents requires high costs. Therefore, computer scientists started to analyze human written texts data by using statistical approaches so that a machine can deal with huge amount of documents instead of human. However, analyzing the human language is more than just statistics. It is much far beyond mathematics than we simply expect. Therefore, scientists fell into deep agony to overcome this issue. One of the possible ideas is building a machine readable knowledge system followed by the human brain system. It was based on the hypothesis that human understands meaning of words by using concepts in his/her memory which he/she have studied before. If a machine can have knowledge system, it might be possible to analyze documents more smartly and precisely than ever. The biggest problem for understanding human language by computer is that a word can have multiple meanings, Word Sense Disambiguation (WSD) problem is the most challenging issue to be solved. Therefore, scientists have focused on the WSD problem. The one of the most popular algorithm based on the machine readable knowledge is Structural Semantic Interconnections (SSI) algorithm which applies the hypothesis that a concept of the given word can be disambiguated by comparing interconnections between concepts of co-occurring words. Even though, the SSI algorithm is a powerful method, it still has a weakness to overcome. A word ambiguity is different from each other words. Some words have only single meaning (monosemous word) but some words have multiple meanings (polysemous word). The word which has low word ambiguity must be disambiguated earlier than the words with high word ambiguity. Moreover, a word is likely to be semantically related with the adjacent words. If the centroid word and the target word to be analyzed are adjacent in the given sentence, it has higher possibility to share semantic relations than the words far apart. In order to apply these two hypotheses, the Low Ambiguity First (LAF) algorithm has been introduced in this research. Word ambiguity will be measure by using a number of possible concepts and frequencies of concepts of target words which are defined in the WordNet. In order to demonstrate the superiority of the proposed algorithm, nouns in SemCor2.1 corpus have been disambiguated by using the base-line, the SSI, and the LAF algorithm. Experimental results clearly shows that the proposed algorithm disambiguates nouns more accurately than other algorithms (Brown1: 9.546% improved, Brown2: 10.324% improved). As a result, the proposed LAF algorithm can disambiguate nouns semantically with the highest precision ratio. However, the weakness of the proposed algorithm is that it is depends on the performance of the WordNet. If target words are not defined in the WordNet, there is no way to disambiguate the target words. Hence, further extension of the LAF algorithm includes additional works to deal with words like proper nouns or technical terms by using web resources such as the Wikipedia.

도메인 지식 구축에 의한 의미적 비디오 이벤트 표현

송단 Chosun Univ. 2005 국내석사

The MPEG-7 visual standard under development specifies content-based descriptors that allow users or agents (or search engines) to measure similarity in images or video based on visual criteria, and can be used to efficiently identify, filter, or browse images or video based on visual content. More specifically, MPEG-7 specifies color, texture, object shape, global motion, or object motion features for this purpose. This paper outlines the aim, methodologies, and broad details of the MPEG-7 standard development for video event description. Except for assistant by the MPEG-7 tools, we also put forward a novel method for video event analysis and description based on the Domain Knowledge in this paper. Semantic concepts in the context of the video event are described in one specific domain enriched with qualitative attributes of the semantic objects, multimedia processing approaches and domain independent factors: low level features (pixel color, motion vectors and spatio-temporal relationship). In order to apply large-scale semantic knowledge in vision problems effectively, catering the naive user’s retrieval and index processing with semantic (human) language, a few major issues are resolved in this paper. Firstly, how can we get the semantic shot for the specific Domain Knowledge? The former existing algorithm has been adopted to solve the problem. Secondly, what visual observables should be collected? This is usually dependent on the problem domain. Here, we consider one shot of the billiard game clip as the specific Domain Knowledge. Thirdly, how can these observables be translated into the semantic representation, we are from two aspects to expose that issue: Firstly, video event representation using MPEG-7 high level descriptors which was defined in the MPEG-7 XML files. Secondly, video object motion analysis with the help of the MPEG-7 low level descriptors(video object motion detection and moving trajectory analysis). In addition, the most important contribution in this work is exploiting the video object ontology to map the MPEG-7's high-level descriptors to low level features descriptors which have been defined in the MPEG's logical structure. 현재 표준안으로 만들어지고 있는 MPEG-7에서는 저차원의 특징뿐만 아니라 시공간적 관계 표현, 이벤트 인식에 이르기까지 의미적 인식을 위한 노력을 하고 있으나 이는 메타 데이터 형태로 미디어객체에 대해 단순히 키워드를 부여하는 방법으로 그 내용을 표현하는 정도이고, 진정한 의미적 내용을 표현하는 것은 현재로서는 불가능하다. 이에 비디오 내용을 의미적으로 인식하기 위해서는, 비디오내 움직임 객체의 섬세하고 세밀한 의미적 표현이 필수적이라고 생각된다. 본 논문에서는 현재까지 기술 개발된 움직임 객체의 분리추적기술 등을 기반으로 객체들 간의 시공간적 관계 표현 및 이들의 표현을 매칭시키기 위한 방법으로 온톨로지를 적용하는 방안을 제시하고자 한다. 이는 추후 멀티미디어 데이터 처리에 많은 응용이 발생할 것으로 보인다.

학교 웹사이트 평가 모형에 관한 연구

정세정 조선대학교 2004 국내석사

The Internet which brought about information revolution and drastic changes to life style is giving a lot of impact on education. The Internet can be an effective self-study tool both at school and at home. The purpose of this paper is to propose an analytical model for evaluating school web sites. It is in the researches, some important aspects to be considered to evaluate web sites, and particulary too unilateral criteria were applied to all aspects to conduct accurate evaluation of school web sites. In order to tell which web sites provide worthful information more effectively, researches have been conducted to evaluate web sites. The criteria consists of seven categories, including design, reliability, interface, technology, community, educational contents. The results show that 'educational contents(26.3%)' is the most important and secondly is 'reliability(19.5%)' for the school web sites. This paper will provide new criteria for planning, creating, modifying or upgrading school web sites. This kind of evaluation will distinguish educationally-effective sites to be used for learning. The evaluation method was designed to be the criteria for planning or creating a website and serve as basis for modifying or upgrading the school web site.

멀티모달 AI를 적용한 웹툰 생성 연구

유경호 조선대학교 대학원 2023 국내박사

In this thesis, I conducted research on generating webtoons using deep learning-based text-to-image generation technology to assist webtoon creators in their creative activities. The research methodology involved constructing a multimodal webtoon dataset by using publicly available datasets such as MSCOCO. The generated dataset consists of treatments (that is text descriptions) and their corresponding webtoon – treatment-webtoon dataset. Furthermore, continuous text data was also collected using ChatGPT. To generate webtoon, this thesis proposed to utilize a multilingual BERT model for feature extraction from the treatments, add noise to them, and input the noisy features into a DCGAN (Deep Convolutional Generative Adversarial Network). The experimental results showed relatively low performance with an inception score of 4.9 and FID (Fréchet Inception Distance) of 22.21. To overcome the limitations of DCGAN, this thesis proposed to train the CLIP (Contrastive Language-Image Pretraining) model on the treatment-webtoon dataset by measuring the similarity between text and images and then use the diffusion model to generate webtoon. CLIP is a model that can learn a relationship between multimodal data (such as the treatment-webtoon dataset) by extracting features from each data modality. The goal is to bring similar data closer and dissimilar data farther apart in the same feature space, which is achieved by contrastive learning. In this thesis, the performance of CLIP trained on the treatment-webtoon dataset was evaluated using quantitative metrics such as measuring similarity between bilingual treatments and images, treatment-based search queries for similar images, and zero-shot classification. For generating webtoons based on the diffusion model, a desired text along with its CLIP features and the image with the most similar CLIP features were inputted into a pretrained depth-to-image model. In the experiments, webtoons were generated using both single text and continuous text inputs. The results showed that when using continuous text inputs for webtoon generation, the inception score improved from 4.9, as in the case of DCGAN, to 7.14 and the generated images were of higher quality. The technology developed in this thesis can be used by webtoon creators by inputting their desired text to generate webtoons more efficiently and in a timely manner. However, one of the main limitations of this work is that currently it cannot generate webtoons from multiple sentences and images or maintain consistent artistic style throughout the generated images. Therefore, further research is needed on diffusion models that can handle multiple sentences as inputs and generate images with consistent artistic styles when continuous text inputs are provided.

지능적인 웹 검색을 위한 의미적 문서 태깅 방법 연구

황명권 조선대학교 대학원 2011 국내박사

Nowadays, the fast advance of digital technologies and the current Web environment have been accelerating the field of information retrieval and processing. The Internet space using the Web is not strange any more to most people and they can obtain any information desired from the Web. These changes have spawned a great deal of research aiming at enhancing service and convenience. Thus, many computer science researchers are committed to finding more useful and efficient methods to provide appropriate results to meet users' needs. Among those, the methods of this dissertation have been studied for semantic document tagging to realize Semantic Web as an ultimate purpose. Semantic Web is a very important technique aiming at processing and understanding the information spread on the Web and subsequently providing semantic and exact retrieval results. To realize Semantic Web, this research concentrates on tagging methods of text documents. The amount of the texts is increasing according to trend of Web 2.0 and it is the most frequently utilized communication medium to express and share information between people. Therefore, the text retrieval is important and this research proposes tagging methods of Web documents to provide standardized, systematic and semantic retrieval. The previous works on Web document tagging generally choose core words from a document itself. However, the core words are not standardized taggers so, in retrieving, users should make an effort to grasp the tagger words first. To improve the point, this research contains methods to utilize titles (Wiki concept) of Wikipedia documents and to find the best Wiki concept which describes the Web documents (target documents). In addition to these methods, the research tries to classify target documents into Wikipedia category (Wiki category) for semantic document interconnections. In order to use Wiki categories and concepts for classifying and tagging target documents, the research extracts context information from Wiki concepts, Wiki categories and target documents and finds the nearest Wiki categories and concepts of target documents through similarity measure. Experimenting diverse cases, it was confirmed that this research can provide semantic classification and tagging methods and that the context information of documents has much potentiality to be applied to various works for Semantic Web. By the way, it is worth noting that some future works, which can give semantics to proper nouns and technical terms, need to be done.