• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2018.06a
• /
• pp.223-224
• /
• 2018

기계학습 기반 비디오 화질 자동 측정 기술은 주관적 화질 평가를 대체하기 위한 기술로, 비디오를 입력 신호로 화질 평가 결과를 출력 신호로 하는 기계학습 모델을 통해서 개발하는 기술이다. 학습에 필요한 비디오 데이터 셋은 입력 신호인 비디오 시퀀스와 입력의 출력신호로 학습할 주관적 화질 평가 결과로 구성된다. 이때 데이터 셋의 일부는 기계학습 기반 비디오 화질 자동 측정 기술 개발 과정에서 학습에 사용하고, 남은 일부는 개발 기술의 성능 평가에 사용한다. 일반적으로 기계학습 기반 기술의 성능은 학습 데이터의 양과 질에 비례한다. 그러나, 기계학습 기반 비디오 화질 자동 측정 기술 개발에 필요한 데이터 셋은 주관적 화질 평가 결과를 포함해야 하므로, 데이터 양을 늘리는 것은 쉬운 문제가 아니다. 이에 본 논문에서는 압축 비디오에 대한 화질 자동 측정 기술 개발을 위해 필요한 데이터 셋을 양과 질적 측면에서 효율적으로 구축하는 방법을 제안한다. 양적 측면에서 효율성을 높이기 위해 부호화 복잡도와 평가 난이도 기반으로 시퀀스를 선정 방법을, 질적 측면에서 효율성을 높이기 위해 쌍 비교(Pairwise Comparison)기반의 주관적 화질 평가 방법을 제안한다.

• Annual Conference on Human and Language Technology
• /
• 2019.10a
• /
• pp.470-474
• /
• 2019

한글 OCR 성능을 높이기 위해 딥러닝 모델을 활용하여 문자인식 부분을 개선하고자 하였다. 본 논문에서는 폰트와 사전데이터를 사용해 딥러닝 모델 학습을 위한 한글 문장 이미지 데이터를 직접 생성해보고 이를 활용해서 한글 문장의 OCR 성능을 높일 다양한 모델 조합들에 대한 실험을 진행했다. 딥러닝 모델은 STR(Scene Text Recognition) 구조를 사용해 변환, 추출, 시퀀스, 예측 모듈 각 24가지 모델 조합을 구성했다. 딥러닝 모델을 활용한 OCR 실험 결과 한글 문장에 적합한 모델조합은 변환 모듈을 사용하고 시퀀스와 예측 모듈에는 BiLSTM과 어텐션을 사용한 모델조합이 다른 모델 조합에 비해 높은 성능을 보였다. 해당 논문에서는 이전 한글 OCR 연구와 비교해 적용 범위를 글자 단위에서 문장 단위로 확장하였고 실제 문서 이미지에서 자주 발견되는 유형의 데이터를 사용해 애플리케이션 적용 가능성을 높이고자 한 부분에 의의가 있다.

• Annual Conference on Human and Language Technology
• /
• 2023.10a
• /
• pp.50-55
• /
• 2023

속성 기반 감정 분석은 텍스트 내에서 감정과 해당 감정이 특정 속성, 예를 들어 제품의 특성이나 서비스의 특징에 어떻게 연결되는지를 분석하는 태스크이다. 본 논문에서는 속성 기반 감정 분석 데이터를 사용한 다중 작업-토큰 레이블링 문제에 프롬프트 튜닝 기법을 적용하기 위한 포괄적인 방법론을 소개한다. 이러한 방법론에는 토큰 레이블링 문제를 시퀀스 레이블링 문제로 일반화하기 위한 감정 표현 영역 검출 파이프라인이 포함된다. 또한 분리된 시퀀스들을 속성과 감정에 대해 분류 하기 위한 템플릿을 선정하고, 데이터셋 특성에 맞는 레이블 워드를 확장하는 방법을 제안함으써 모델의 성능을 최적화한다. 최종적으로, 퓨샷 세팅에서의 속성 기반 감정 분석 태스크에 대한 몇 가지 실험 결과와 분석을 제공한다. 구축된 데이터와 베이스라인 모델은 AIHUB(www.aihub.or.kr)에 공개되어 있다.

• The Journal of Korean Institute of Electromagnetic Engineering and Science
• /
• v.14 no.12
• /
• pp.1239-1247
• /
• 2003

OFDM(orthogonal frequency division multiplexing) communications system is very attractive for the high data rate transmission in the frequency selective lading channel. Since OFDM has high PAPR(peak-to-average power ratio), OFDM signal may be distorted by the nonlinear HPA(high power amplifier). In this paper, we propose the DSI(dummy sequence insertion) method for OFDM communication system. Some sub-carriers are inserted for PAPR reduction. They carry the specified dummy data sequence which are used for only PAPR reduction and do not work as side information like the conventional PTS(partial transmit sequence) or SLM(selected mapping) method. We use the complementary sequence and the combination of the correlation sequence as the dummy sequence. Flipping technique is used for the DSI method to get the effective PAPR reduction. It is important that BER of the proposed method is independent of the damage of the dummy data sequence. And DSI method has better spectral efficiency than the conventional block coding. On the other hand, threshold PAPR method is applied to cut down the processing time. However, this DSI method is not better than the conventional PTS method in the respect of the PAPR reduction performance. The DSI method includes the threshold PAPR lower than the PAPR of the OFDM signal, reduces the processing time and improves the BER performance.

• The KIPS Transactions:PartD
• /
• v.16D no.2
• /
• pp.237-248
• /
• 2009

Recently, based on dynamic location or mobility of moving object, many researches on pattern mining methods actively progress to extract more available patterns from various moving patterns for development of location based services. The performance of moving pattern mining depend on how analyze and process the huge set of spatio-temporal data. Some of traditional spatio-temporal pattern mining methods[1-6,8-11]have proposed to solve these problem, but they did not solve properly to reduce mining execution time and minimize required memory space. Therefore, in this paper, we propose new spatio-temporal pattern mining method which extract the sequential and periodic frequent moving patterns efficiently from the huge set of spatio-temporal moving data. The proposed method reduces mining execution time of $83%{\sim}93%$ rate on frequent moving patterns mining using the moving sequence tree which generated from historical data of moving objects based on hash tree. And also, for minimizing the required memory space, it generalize the detained historical data including spatio-temporal attributes into the real world scope of space and time using spatio-temporal concept hierarchy.

• Journal of Applied Biological Chemistry
• /
• v.59 no.4
• /
• pp.299-304
• /
• 2016

When used for amplicon sequencing, Illumina platforms produce more than hundreds of sequence artefacts, which affects operational taxonomic units based analyses such as differential abundance and network analyses. Nevertheless it has become a major tool for fecal microbial community analysis. In addition, results from sequence-based fecal microbial community analysis vary depending on conditions of samples (i.e., freshness, time of storage and quantity). We investigated if freeze-drying samples could improve quality of sequence data. Our results showed reduced number of possible artefacts while maintaining overall microbial community structure. Therefore, freeze-drying feces prior to DNA extraction is recommended for Illumina-based microbial community analysis.

• Journal of KIISE:Databases
• /
• v.29 no.5
• /
• pp.381-392
• /
• 2002

This paper deals with the problem of shape-based retrieval in time-series databases. The shape-based retrieval is defined as the operation that searches for the (sub)sequences whose shapes are similar to that of a given query sequence regardless of their actual element values. In this paper, we propose an effective and efficient approach for shape-based retrieval of subsequences. We first introduce a new similarity model for shape-based retrieval that supports various combinations of transformations such as shifting, scaling, moving average, and time warping. For efficient processing of the shape-based retrieval based on the similarity model, we also propose the indexing and query processing methods. To verify the superiority of our approach, we perform extensive experiments with the real-world S&P 500 stock data. The results reveal that our approach successfully finds all the subsequences that have the shapes similar to that of the query sequence, and also achieves significant speedup up to around 66 times compared with the sequential scan method.

• The Transactions of the Korea Information Processing Society
• /
• v.6 no.4
• /
• pp.868-876
• /
• 1999

In this paper, an efficient automatic video parsing technique on MPEG-compressed video that is fundamental for content-based indexing is described. The proposed method detects scene changes, regardless of IPB picture composition. To detect abrupt changes, the difference measure based on the dc coefficient in I picture and the macroblock reference feature in P and B pictures are utilized. For gradual scene changes, we use the macroblock reference information in P and B pictures. the process of scene change detection can be efficiently handled by extracting necessary data without full decoding of MPEG sequence. The performance of the proposed algorithm is analyzed based on precision and recall. the experimental results verified the effectiveness of the method for detecting scene changes of various MPEG sequences.

• Journal of the Institute of Electronics Engineers of Korea TC
• /
• v.45 no.2
• /
• pp.6-12
• /
• 2008

In uplink cable modem link operated in time-division multiple access mode, it is crucial to employ a suitable preamble pattern enabling frame detection, coarse timing/carrier recovery. Preamble pattern based on the constant envelope zero autocorrelation sequence is proposed for the uplink cable modem compliant to the data over cable service interface specification. Frame detection, coarse/fine timing and carrier recovery algorithms suitable for the proposed preamble pattern are also proposed. We check up the performances using numerical results.

• Journal of the Institute of Convergence Signal Processing
• /
• v.10 no.2
• /
• pp.132-135
• /
• 2009

The video similarity measure is very important factor to index and to retrieve for video data. In this paper, we propose the luminance projection model to measure the video similarity efficiently. Most algorithms for video indexing have been commonly used histograms, edges, or motion features, whereas in this paper, the proposed algorithm is employed an efficient measure using the luminance projection. To index effectively the video sequences and to decrease the computational complexity, we calculate video similarity using the key frames extracted by the cumulative measure, and compare the set of key frames using the modified Hausdorff distance. Experimental results show that the proposed luminance projection model yields the remarkable accuracy and performance than the conventional algorithm.