• ETRI Journal
• /
• 제18권1호
• /
• pp.29-41
• /
• 1996

This paper describes a method to select a suitable feature for speech recognition using information theoretic measure. Conventional speech recognition systems heuristically choose a portion of frequency components, cepstrum, mel-cepstrum, energy, and their time differences of speech waveforms as their speech features. However, these systems never have good performance if the selected features are not suitable for speech recognition. Since the recognition rate is the only performance measure of speech recognition system, it is hard to judge how suitable the selected feature is. To solve this problem, it is essential to analyze the feature itself, and measure how good the feature itself is. Good speech features should contain all of the class-related information and as small amount of the class-irrelevant variation as possible. In this paper, we suggest a method to measure the class-related information and the amount of the class-irrelevant variation based on the Shannon's information theory. Using this method, we compare the mel-scaled FFT, cepstrum, mel-cepstrum, and wavelet features of the TIMIT speech data. The result shows that, among these features, the mel-scaled FFT is the best feature for speech recognition based on the proposed measure.

• Journal of Advanced Marine Engineering and Technology
• /
• 제36권5호
• /
• pp.662-673
• /
• 2012

This study introduces a condition diagnosis technique for a turbine governor system. The governor system is an important control system to handle turbine speed in a nuclear power plant. The turbine governor system includes turbine valves and stop valves which have their own functions in the system. Because a turbine governor system is operated by high oil pressure, it is very difficult to maintain under stable operating conditions. Turbine valves supply oil pressure to the governor system for proper operation. Using the pressure variation of turbine and governor valves, operating conditions of the turbine governor control system are detected and identified. To achieve automatic detection of valve status, time-based and frequency-based analysis is employed. In this study, a new approach, wavelet decomposition, was used to extract specific features from the pressure signals of the governor and stop valves. The extracted features, which represent the operating conditions of the turbine governor system, include important information to control and diagnose the valves. After extracting the specific features, decision rules were used to classify the valve conditions. The rules were generated by a decision tree algorithm (a typical simple method for data-based rule generation). The results given by the wavelet-based analysis were compared to detection results using time- and frequency-based approaches. Compared with the several related studies, the wavelet transform-based analysis, the proposed in this study has the advantage of easier application without auxiliary features.

• 방송공학회논문지
• /
• 제25권1호
• /
• pp.58-66
• /
• 2020

영상 촬영 시 객체의 움직임, 탈초점(Out-of-focus) 등의 이유로 영상 흐려짐 현상이 빈번하게 발생하며, 이 과정에서 선명한 영역의 고주파 성분이 급격하게 감소하게 된다. 이러한 성질을 바탕으로, 본 논문에서는 질감 특징 표현자를 사용하여 별도의 주파수 변환 과정 없이 고주파 성분을 추정하고, 이를 바탕으로 흐려진 영역을 검출하는 방법을 제안한다. 제안하는 방법은 학습 기반 질감 표현자와 유역(Watershed) 기반 질감 표현자를 함께 이용하여 다양한 환경에서도 흐려진 영역을 검출할 수 있다. 또한, 흐려짐을 검출하는 최소 단위를 화소 단위에서 영역 단위로 확장하여 처리 속도를 향상시키고, 영상 보정 기법을 이용하여 흐려짐 검출 성능을 개선하였다. 실험 결과는 제안하는 방법이 기존의 흐려짐 검출 방법 대비 성능이 향상되었음을 보여준다.

• 인터넷정보학회논문지
• /
• 제10권5호
• /
• pp.127-133
• /
• 2009

본 논문은 가중 퍼지소속함수 기반 신경망(Neural Network with Weighted Fuzzy Membership Functions, NEWFM)을 이용하여 간질 증세를 가진 사람과 건강한 사람의 뇌파(electroencephalogram, EEG)로부터 정상 파형과 간질(epilepsy) 파형을 분류하는 방안을 제시하고 있다. NEWFM에서 사용할 특징입력을 추출하기 위해서 첫 번째 단계에서는 웨이블릿 변환(wavelet transform, WT)을 이용하였다. 두 번째 단계에서는 첫 번째 단계에서 생성한 웨이블릿 계수들을 주파수 분포와 주파수 변동량을 이용하여 24개의 특징입력을 추출하였다. NEWFM은 이들 24개의 특징입력을 이용하여 정상 파형과 간질 파형을 분류하였을 때 98%의 분류성능을 나타내었다.

• 로봇학회논문지
• /
• 제1권1호
• /
• pp.25-35
• /
• 2006

Improving practicality of SLAM requires various sensors to be fused effectively in order to cope with uncertainty induced from both environment and sensors. In this case, combining sonar and vision sensors possesses numerous advantages of economical efficiency and complementary cooperation. Especially, it can remedy false data association and divergence problem of sonar sensors, and overcome low frequency SLAM update caused by computational burden and weakness in illumination changes of vision sensors. In this paper, we propose a SLAM method to join sonar sensors and stereo camera together. It consists of two schemes, extracting robust point and line features from sonar data and recognizing planar visual objects using multi-scale Harris corner detector and its SIFT descriptor from pre-constructed object database. And fusing sonar features and visual objects through EKF-SLAM can give correct data association via object recognition and high frequency update via sonar features. As a result, it can increase robustness and accuracy of SLAM in indoor environment. The performance of the proposed algorithm was verified by experiments in home -like environment.

• ETRI Journal
• /
• 제43권1호
• /
• pp.95-108
• /
• 2021

Application (app) ratings are feedback provided voluntarily by users and serve as important evaluation criteria for apps. However, these ratings can often be biased owing to insufficient or missing votes. Additionally, significant differences have been observed between numeric ratings and user reviews. This study aims to predict the numeric ratings of Google apps using machine learning classifiers. It exploits numeric app ratings provided by users as training data and returns authentic mobile app ratings by analyzing user reviews. An ensemble learning model is proposed for this purpose that considers term frequency/inverse document frequency (TF/IDF) features. Three TF/IDF features, including unigrams, bigrams, and trigrams, were used. The dataset was scraped from the Google Play store, extracting data from 14 different app categories. Biased and unbiased user ratings were discriminated using TextBlob analysis to formulate the ground truth, from which the classifier prediction accuracy was then evaluated. The results demonstrate the high potential for machine learning-based classifiers to predict authentic numeric ratings based on actual user reviews.

• 한국가구학회지
• /
• 제16권1호
• /
• pp.17-23
• /
• 2005

European maple is famous for the optimum solid wood for making bridge which is the most important part in violin acoustics. This study was carried out to investigate the variation of main features, i. e. annual ring width and cutting direction of costly imported violin bridge blanks and to examine the effect of these features of the blanks on the vibrational characteristics of bridge blanks. Imported violin bridge blanks had somewhat large variation in major macroscopical and physical properties and there was little relationship between annual ring density and weight of maple blanks. Resonant frequency of violin bridge blanks had some positive correlation with weight, however, damping having negative relationship with frequency was seldom affected by any physical properties of the maple blanks. Deviation from the radial cutting of tail side(ray direction from top toward feet on the edge of bridge blank) lowered the resonant frequency. Consequently, weight and ray direction should be taken for the critical quality decisive factors(QDF) of incoming bridge blanks by not only inspectors also luthiers who tune the bridge by shaping and are responsible for the final timbre quality of this complicate instrument.

• 말소리와 음성과학
• /
• 제2권2호
• /
• pp.51-60
• /
• 2010

This study is to examine how Korean speakers realize English stress on the homographic words. Experiments were performed by Korean speakers three times, before stress instruction, immediately after instruction, and six weeks after instruction. First, duration, fundamental frequency, and intensity of the vowel in a stressed syllable of three homographic words produced by Korean speakers were compared with those of native speakers of English. The result shows that when the words were used as nouns, before instruction Korean speakers had shorter duration and lower fundamental frequency in the stressed vowel than the native speakers, which indicates that Korean speakers did not assign the primary stress on the first syllable of the nouns. After instruction, the values of duration and fundamental frequency were increased and the differences between two groups were decreased. Next, the values of these stress features measured three times were analyzed in order to find out how they changed through instruction. The analysis shows that after instruction the values of three features were increased compared to the ones before instruction, and that the biggest change was in duration of the vowel and the next was fundamental frequency. Six weeks after instruction, the values of duration and intensity were decreased than those immediately after instruction. This means that instruction is helpful for Korean speakers to assign the stress for the English homographic words, and that instruction and practice are needed repeatedly.

• 감성과학
• /
• 제10권3호
• /
• pp.373-381
• /
• 2007

본 연구의 목적은 치매노인의 보호를 위해 필요한 주요문제특성의 추출이다. 시설에 입소해 있는 치매노인 110명과 정상노인 30명을 대상으로 기본특성 조사, 인지기능 선별검사(Korean Mini-Mental State Examination: K-MMSE), 문제특성 설문조사를 실시하였다. 요인분석(factor analysis)을 이용하여 치매노인의 주요문제특성을 추출하고, 일원변량분석(one-way ANOVA)을 이용하여 성별과 치매 중증도에 따른 문제특성의 빈도차이를 검증하였다. 요인분석결과 총 20개의 주요문제특성이 추출되었다. 문제특성의 빈도는 성별에 따라 폭력적 언어 및 작화, 수집행동, 반복행동에서 유의한 차이가 있었다. 치매 중증도에 따라서는 성적 이상행동과 시청력 장애를 제외한 모든 특성에서 유의한 차이가 나타났다. 본 연구 결과는 치매노인 보호를 위한 유비쿼터스 기술의 활용에 기초자료를 제공할 수 있을 것이다.

• 정보처리학회논문지B
• /
• 제13B권4호
• /
• pp.489-498
• /
• 2006

색인전문가에 의해 분류된 웹문서들을 통계적 자질 선택방법으로 자질을 추출하여 클라스터링을 해 보면, 자질 선택에 사용된 데이터셋에 따라 성능과 결과가 다르게 나타난다. 그 이유는 많은 웹 문서에서 문서의 내용과 관계없는 단어들을 많이 포함하고 있어 문서의 특정을 나타내는 단어들이 상대적으로 잘 두드러지지 않기 때문이다. 따라서 클러스터링 성능을 향상시키기 위해 이런 부적절한 자질들을 제거해 주어야 한다. 따라서 본 논문에서는 자질 선택에서 자질의 문서군별 자질값뿐만 아니라, 문서군별 자질값의 분포와 정도, 자질의 출현여부와 빈도를 고려한 자질 필터링 알고리즘을 제시한다. 알고리즘에는 (1) 단위 문서 내 자질 필터링 알고리즘(FFID : feature filtering algorithm in a document), (2) 전체 데이터셋 내 자질 필터링 알고리즘(FFIM : feature filtering algorithm in a document matrix), (3)FFID와 FFIM을 결합한 방법(HFF:a hybrid method combining both FFID and FFIM) 을 제시한다. 실험은 단어반도를 이용한 자질선택 방법, 문서간 동시-링크 정보의 자질확장, 그리고 위에서 제시한 3가지 자질 필터링 방법을 사용하여 클러스터링 했다. 실험 결과는 데이터셋에 따라 조금씩 차이가 나지만, FFID보다 FFIM의 성능이 좋았고, 또 FFID와 FFIM을 결합한 HFF 결과가 더 나은 성능을 보였다.