Abstract:
PURPOSE: A system and a method for handling occlusion in stereo imaging, and a computer-readable medium for recording an occlusion handling algorithm are provided to display the shape of an object clearly by detecting, compensating and matching an occlusion region and a solid color region. CONSTITUTION: A parallax estimation unit(140) compares pixels within an over-segment of over-split reference and response image to the parallax between pixels, and compares the pixels within the under-segment of a low-split reference and response image to estimate the parallax between the pixels. An occlusion detection unit(150) warps the pixels of the reference and response images based on the parallax in order to detect occlusion pixels which are not matched. A stereoscopic matching unit(160) implements energy terms for processing the occlusion pixels and a solid color region in order to perform stereoscopic matching.
Abstract:
An apparatus for constructing an acoustic model on the basis of the Hidden Markov model and a method thereof are provided to construct more reliable a voice recognition system of which recognition is improved by using an erroneous restriction condition while using a maximal mutual information scheme. An acoustic feature extracting unit(200) extracts training voice features, which are currently used, to constructing an acoustic mode based on the Hidden Markov. An acoustic model initialization setting unit(202) determines the initial value of a parameter of the acoustic mode to maximize the log likelihood of the training voice for the predetermined correct answer model. When the training voice is recognized by using the acoustic model having the initial value calculated by the acoustic model initialization setting unit, a correct answer model log likelihood calculating unit(206) calculates the log likelihood as to a training voice correct answer model based on a parameter of the acoustic model.
Abstract:
A method for eliminating aliasing in a sub band adaptive filtering system and the sub band adaptive filtering system without alias are provided to improve the performance of the sub band adaptive filtering system by removing an interband aliasing extracted from each sub band signal by using a bandwidth-increased linear phase FIR analysis filter. An input signal is passed through a plurality of sub-band analysis filters to remove the input signal passed through the unknown system. The passed input signal is classified into a plurality of sub-band signals. The sub-band signal is down-sampled. The down sampled sub band signal is passed through the corresponding sub band adaptive filter. An interband aliasing signal is extracted from the down sampled sub band signal. The interband aliasing signal extracted from the down sampled sub band signal is removed. An aliasing extraction filter extracts the interband aliasing signal from the down sampled sub band signal.
Abstract:
An apparatus and a method for recognizing a speaker are provided to improve the accuracy in recognition of the speaker by using the reliabilities of speaker voices, which are received in the past. A reference voice database stores reference voices, which correspond to voices of one or more users registered, while the reference voices are matched to the users. A voice reception unit receives a voice of a speaker. A voice feature extraction unit extracts one or more features for speaker-recognition from the received voice. A reliability measurement unit measures the similarity between the extracted features of the received voice and each of the reference voices, thereby obtaining the reliability of voice, which represents a possibility that the speaker corresponds each of the users. A reliability database stores the measured reliability for each of the users. A weight computation unit computes respective weight values for a series of reliabilities stored in the reliability database, according to a predetermined rule. A weight appliance unit applies the computed weight values to the reliabilities, respectively. A determination unit determines whether the received voice is identical to one of the reference voices by calculating an average of the reliabilities to which the weight values are respectively applied and comparing the average with the reliability of the received voice.
Abstract:
본 발명은 정규화된 스펙트럼 부밴드 중심점(Normalized Spectral Subband Centroid; NSSC)을 기반으로 핑거프린트를 생성하는 방법과, 미지의 오디오 신호가 입력으로 주어졌을 때, 이를 이미 구축되어 있는 대용량 오디오 데이터베이스에서 검색하여 입력 오디오 신호에 대한 정보를 출력해 주기 위한 오디오 핑거프린팅 시스템에 대한 것이다. 오디오 핑거프린팅 시스템은 사용된 핑거프린트에 의해 그 성능이 크게 좌우된다. 본 발명의 NSSC 핑거프린트는 오디오 신호의 특징을 잘 나타내어 인식에 사용하기 적합하면서도 오디오 신호에 가해질 수 있는 여러 가지 왜곡들, 예를 들어 MP3 압축, 이퀄라이제이션(equalization) 등에 매우 강인하며, 대용량 데이터베이스 구축과 실시간 검색에도 유리한 장점을 가지고 있다. 실험 결과에 따르면, 본 발명에 의한 시스템은 기존의 오디오 핑거프린팅 시스템에 비해 향상된 성능을 보인다. 본 발명은 인터넷 상의 오디오 불법 유통을 막기 위한 실시간 필터링 서비스, 대용량 오디오 데이터베이스의 자동 인덱싱(indexing), 그리고 방송 모니터링 등에 응용될 수 있다. 핑거프린트, 정규화된 스펙트럼 부밴드 중심점, 오디오, MP3
Abstract:
행동패턴 분석 장치 및 방법이 개시된다. 행동패턴 분석 장치는 영상 촬영 장치를 통해 수집되는 영상 데이터를 입력받는 영상 입력부, 영상 데이터로부터 관심 객체를 검출하기 위하여 일정 단위 시간 동안 입력된 영상 데이터를 이용하여 배경 모델링을 수행하는 배경 모델링부, 배경 모델링을 통해 학습된 배경 모델을 이용하여 관심 객체를 검출하는 객체 검출부, 영상 데이터로부터 관심 객체의 행동 패턴에 대한 특징을 추출하는 특징 추출부, 추출된 특징을 이용하여 관심 객체의 행동 패턴을 학습하여 모델링하는 행동패턴 모델링부 및 모델링된 행동 패턴을 분석하여 관심 객체의 특이 행동 이벤트 발생 여부를 판단하는 분석부를 포함한다.