-
公开(公告)号：KR101359689B1
公开(公告)日：2014-02-10
申请号：KR1020120006898
申请日：2012-01-20
Applicant: 한국과학기술원
CPC classification number: G10L15/148 , G10L2015/025
Abstract: 음성 인식 시스템에서 음소를 인식하는 방법은 음소 데이터 인식 장치에서 음성을 수신하는 단계; 및 음소 데이터 처리 장치에서 수신된 상기 음성을 세그먼트 기반의 준-마르코프 모델을 이용하여 음소를 인식하는 단계를 포함한다.
-
公开(公告)号：KR1020130085698A
公开(公告)日：2013-07-30
申请号：KR1020120006669
申请日：2012-01-20
Applicant: 한국과학기술원
IPC: G06T7/00
Abstract: PURPOSE: An image partitioning method using correlation clustering, a system processing the same, and a recording medium are provided to eliminate the need of predetermining the number of partition region or partitioning degree of each image, thereby implementing image partitioning customized for a task through training. CONSTITUTION: A system divides an input image into super pixels (S210). The system constructs a graph from two neighboring super pixels among the obtained super pixels through a connection in consideration of a specific condition (S220). The system extracts a feature vector of each edge of the graph, forms a joint feature map, and partitions the constructed graph through correlation clustering (S230). [Reference numerals] (AA) Start; (BB) End; (S210) System divides an input image into super pixels; (S220) System constructs a graph from two neighboring super pixels among the obtained super pixels through a connection in consideration of a specific condition; (S230) System extracts a feature vector of each edge of the graph, forms a joint feature map, and partitions the constructed graph through correlation clustering
Abstract translation: 目的：提供使用相关聚类，系统处理相同的图像分割方法和记录介质，以消除预先确定每个图像的分区数量或分割度的需要，从而通过培训实现为任务定制的图像分区 。 构成：系统将输入图像划分为超像素（S210）。 考虑到特定条件，系统通过连接从所获得的超像素中的两个相邻超像素构成图（S220）。 系统提取图的每个边缘的特征向量，形成联合特征图，并通过相关聚类分割构造的图（S230）。 （附图标记）（AA）开始; （BB）结束; （S210）系统将输入图像划分为超像素; （S220）系统通过考虑到特定条件的连接来构造所获得的超像素中的两个相邻超像素的图; （S230）系统提取图的每个边缘的特征向量，形成联合特征图，并通过相关聚类分割构造的图
-
公开(公告)号：KR101260285B1
公开(公告)日：2013-05-03
申请号：KR1020060117773
申请日：2006-11-27
Applicant: 한국과학기술원
Abstract: 본발명은다원화된확률모형에기반한산술복호화방법에관한것이다. 본발명은특히지상파 DMB 서비스의오디오규격인 BSAC 복호화의핵심이되는산술복호화방법에있어서심볼값 복원을위해확률모형으로부터확률값을추출하는경우, 대규모의일원화된테이블로부터추출하는대신, 다원화된복수의소규모테이블들로부터추출함으로써, 필요한연산량을현저히감소시키는효과를달성할수 있다. 또한본 발명에의한산술복호화방법을디지털신호처리프로세서에의하여구현하는경우, 빠른접근속도와메모리사용의효율적인관리를실현할수 있다.
-
64.发明授权
公开(公告)号：KR101116236B1
公开(公告)日：2012-03-09
申请号：KR1020090069471
申请日：2009-07-29
Applicant: 한국과학기술원
Abstract: 본발명은 WTM(Watson-Tellegen Emotional Model)을기반으로하고 HMM(Hidden Markov Model)과훈련데이타를통한음성의감정에대한훈련을통해서음성에실린감정을인식하는모델을구축하는방법에있어서, WTM의감정군들사이의기하학적거리를사용하여각 감정사이의차이를수치화하는제1단계와제1단계에서설정한값들을기초로하여손실함수(loss function)의값을구하는제2단계및 제2단계에서구해진손실함수를기초로하여각 음성감정모델의파라미터를구하는제3단계를포함하는 WTM을기반으로손실함수와최대마진기법을통한음성감정인식모델구축방법으로써, 이방법을통해구축된음성감정인식모델을통해음성감정인식의성능의향상을기대할수 있다.
-
公开(公告)号：KR101063139B1
公开(公告)日：2011-09-07
申请号：KR1020080093141
申请日：2008-09-23
Applicant: 한국과학기술원
IPC: G06F17/30
Abstract: 본 발명은 영상 자료를 데이터베이스에 입력할 때 영상자료를 분석하여 그 특징을 추출하고, 이를 이용하여 데이터베이스에 인덱스로 사용함으로써 영상 자료 자체를 검색어로 사용할 수 있도록 한 내용기반 영상 검색 방법 장치에 관한 것이다.
이러한 본 발명에 따른 영상 검색은 영역별로 핑거프린트를 추출하여 데이터베이스에 인덱스로 사용함으로써 각 영상의 내용을 기반으로 한 이미지 검색이 가능하고, 특히, 영역 탐색을 통해 추출되는 특정 영역이 영상의 크기 변화나 회전, 절삭에도 왜곡되지 않으므로 검색의 정확도가 높다.
내용기반, 영상 검색, 회전, 절삭, 핑거프린트-
Patent Agency Ranking
公开(公告)号：KR1020110094490A
公开(公告)日：2011-08-24
申请号：KR1020100013892
申请日：2010-02-16
Applicant: 한국과학기술원
CPC classification number: G06F17/30495 , G06F17/16
Abstract: PURPOSE: A higher-order electronic magazine data searching device and method thereof are provided to improve accuracy and reduce calculation amount. CONSTITUTION: A real number vector extractor(300) extracts a real number vector by analyzing input information. A binary part(310) is applied to the quantum scheme and converts the real number vector into a binary data. A candidate generating unit(320) generates a candidate point based on the reliability. A searching unit selects the closest data based on non-similarity between the candidate point and the binary data. The candidate generating unit generates the candidate point by selecting the binary data based on the reliability per bit about the binary data.
Abstract translation: 目的：提供高级电子杂志数据搜索装置及其方法，以提高精度并减少计算量。 构成：实数向量提取器（300）通过分析输入信息提取实数向量。 将二进制部分（310）应用于量子方案，并将实数向量转换为二进制数据。 候选生成单元（320）基于可靠性生成候补点。 搜索单元基于候选点和二进制数据之间的非相似性来选择最接近的数据。 候选生成单元通过基于关于二进制数据的每位的可靠性选择二进制数据来生成候选点。
-
67.发明公开여러 음을 가진 오디오 신호에서 하모닉 구조 모델과 유동적인 길이를 갖는 분석 창을 이용한 멜로디 추출 방법 및 시스템 失效通过使用谐波结构模型和可变长度窗口的方法提取多波音频的装置和方法
公开(公告)号：KR1020110085089A
公开(公告)日：2011-07-27
申请号：KR1020100004695
申请日：2010-01-19
Applicant: 한국과학기술원
IPC: G11B20/10 , G11B27/031
Abstract: PURPOSE: A melody extracting method and system for the same are provided to extract melody component form music with many sounds by using characteristics with energy of the melody component and enabling a melody component to have a harmonic component. CONSTITUTION: A frequency is analyzed by using an analyzing window having a flexible length if an audio signal with many sound is inputted(S100). A candidate melody pitches extracted from an audio signal with a many sounds through a frequency analyzing step by using a harmonic structure model(S200). A melody line is connected to the extracted melody pitch candidates and the candidates area determined(S300).
Abstract translation: 目的：提供一种旋律提取方法和系统，以通过使用具有旋律分量的能量的特征来提取具有许多声音的音乐的旋律分量，并使旋律分量具有谐波分量。 构成：如果输入了许多声音的音频信号，则通过使用具有灵活长度的分析窗口来分析频率（S100）。 通过使用谐波结构模型，通过频率分析步骤从具有许多声音的音频信号提取的候选旋律间距（S200）。 旋律线连接到提取的旋律音调候选和确定的候选区域（S300）。
-
公开(公告)号：KR1020110080554A
公开(公告)日：2011-07-13
申请号：KR1020100000841
申请日：2010-01-06
Applicant: 한국과학기술원
IPC: G11B20/10
CPC classification number: G10L25/90
Abstract: PURPOSE: A method for melody pitch candidate extraction of polyphonic audio based on a harmonic structure model is provided to extract melody pitch candidates in an audio signal including various sounds. CONSTITUTION: An audio signal including various sounds are short Fourier transformed(S10). The unit of the Fourier transformed audio signal is changed into a cent from Hz(S20). A peak point is extracted from each frame of the audio signal which was transformed to the cent unit(S30). The weight of a harmonic structure model which is included in candidate melody pitches is measured(S40). An energy level difference by a frequency is compensated based on the weight information(S50).
Abstract translation: 目的：提供一种基于谐波结构模型的和弦音频旋律音调候选提取方法，以提取包含各种声音的音频信号中的旋律音调候选。 构成：包括各种声音的音频信号是短傅里叶变换（S10）。 傅立叶变换音频信号的单位从Hz变为一分（S20）。 从转换成分单位的音频信号的每个帧提取峰值点（S30）。 测量包括在候选旋律间距中的谐波结构模型的重量（S40）。 基于权重信息来补偿频率的能级差（S50）。
-
公开(公告)号：KR1020110080553A
公开(公告)日：2011-07-13
申请号：KR1020100000840
申请日：2010-01-06
Applicant: 한국과학기술원
IPC: G10K15/00
CPC classification number: G10H1/0025 , G10H2210/056
Abstract: PURPOSE: A melody line decision method from melody pitch candidates on the basis of the property of a melody line is provided for N melody pitch candidates and each melody pitch candidate to efficiently determine a melody line on the basis of physical properties of the melody line. CONSTITUTION: A system is input information about N melody pitch candidates and weight which the melody pitch candidate has in each frame(S10). In a start frame, the system selects N melody pitch candidates and arrays the candidate in a weight descending order(S40). The system determines whether two neighbor melody pitch candidates satisfy set melody line standards(S50). The system selects N melody line candidates through a melody line connection in each melody section(S60). The system selects an optimal melody line(S70).
Abstract translation: 目的：根据旋律线的属性提供旋律音调候选人的旋律线决策方法，用于N个旋律音调候选和每个旋律音调候选，以根据旋律线的物理特性来有效地确定旋律线。 构成：系统是关于N个旋律音调候选的输入信息和旋律音调候选在每个帧中的权重（S10）。 在起始帧中，系统以加权降序选择N个旋律音调候选和阵列候选（S40）。 系统确定两个相邻旋律音调候选是否满足设定的旋律线标准（S50）。 系统通过每个旋律部分中的旋律线连接来选择N个旋律线候选（S60）。 系统选择最佳旋律线（S70）。
-
公开(公告)号：KR1020110036407A
公开(公告)日：2011-04-07
申请号：KR1020090094063
申请日：2009-10-01
Applicant: 한국과학기술원
CPC classification number: G06F17/3002 , G06F17/30156 , G06F17/30306 , G06K9/0002
Abstract: PURPOSE: A method for efficiently learning a distance metric through the minimization of a cost function is provided to improve fingerprint recognition performance by learning a distance metric from training data. CONSTITUTION: A training data is composed of the fingerprint of original content and the fingerprint of distorted content. A distance metric is determined through learning in order to improve recognition performance in the use of training data(S200). A parameterized distance metric is generated in a determination procedure of the distance metric. A cost function is generated. The parameter of the distance metric is determined through the minimization of the cost function.
Abstract translation: 目的：提供一种通过最小化成本函数有效学习距离度量的方法，以通过从训练数据学习距离度量来改善指纹识别性能。 构成：训练数据由原始内容的指纹和失真内容的指纹组成。 通过学习确定距离度量，以便提高训练数据的使用中的识别性能（S200）。 在距离度量的确定过程中生成参数化距离度量。 生成成本函数。 通过成本函数的最小化来确定距离度量的参数。
-
-
-
-
-
-
-
-
-
Search Results
73
records
(took 0.03 seconds)
