고유얼굴을 사용한 검증 기반의 얼굴 추적

정수웅 김천과학대학 2002 김천과학대학 논문집 Vol.28 No.-

얼굴 영상에 대한 처리가 영상 분석 분야에서 많은 관심을 보이고 있다. 본 논문에서는 배경이 복잡한 환경에서 실시간으로 사람의 얼굴을 추적하는 검증 기반 방법을 제안한다. 제안한 방법은 얼굴 추출, 얼굴 검증, 얼굴 추적 과정의 세 단계로 구성되어 있다. 얼굴 추출 과정에서 피부색 모델과 모션 정보를 사용하여 얼굴 후보 영역들을 추출하고 추출된 얼굴 후보 영역이 얼굴인지를 주성분 분석을 통하여 얼굴 검증 과정에서 판별한다. 검증된 얼굴 영역이 화면의 중앙에 항상 위치하도록 능동 카메라를 제어하여 얼굴을 추적한다. 실험을 통하여 제안한 방법이 매우 효과적으로 얼굴을 추적함을 보인다. 제안된 방법은 무인 감시나 사람과 컴퓨터간의 인터페이스 등과 같은 시스템에 응용할 수 있다. The processing of face images receives more and more interesting in the field of image analysis. This paper presents a realtime verification-based face tracking method. The proposed method consists of face detection, verification and tracking modules. First, It detects face candidate regions and verifies candidate regions using principle component analysis. Then, active camera tracks a verified face. Face detection step extracts face candidate regions using skin color and motion information. To improve the robust results of face tracking, the detected regions are verified by eigenfaces. The tracking of a face is done in such a way that the verified region is kept on the center of the screen through controlling an active camera mounted on the pan/tilt platform. The proposed method can be applied to other systems such as intruder inspection system and human-machine interface.

획 해석에 의한 연속 필기 숫자의 On-line 인식에 관한 연구

정수웅 김천과학대학 1994 김천과학대학 논문집 Vol.20 No.-

This paper illustrates an on-line recognition procedure of hand-written number. This recognition procedure is composed of stroke analysis, stroke recognition table, connection between strokes, and number recognition table. The stroke analysis performed the four steps such as partitioning of 4substroke, X axis, Y axis decrement and increment of each substroke, feature point acquisition. In the course of recognizing the stroke, First the feature points of an inputed stroke is acquired through two phases: the first phases defining starting point, end point, and bending point as feature points. Second, using fuzzy 8-direction We investigate the direction componet among feature points. Meanwhile, in number recognition the former stroke and the latter stroke from the set of input strokes is related. And the relative position and adjacency between the former stroke and the latter stroke is considered. Finally, According to the above three form(stroke code, relative position, adjacency) We recognize a handwritten number using number recognition table.

HPC상에서 온라인 문자 인식 시스템

정수웅 김천과학대학 1997 김천과학대학 논문집 Vol.23 No.-

While significant advances have been made in recent years to improve handwritting recognition systems and Hand held PC(HPC), Korean character recognition systems on HPC have still not found broad acceptance in everyday life. In this paper, we present a recognition system of on-line cursive Korean character on HPC. Our system is constructed with hidden Markov model-based character recognition. Hidden Markov model is constructed with nodes for graphemes and edges for Korean combination rules.

스마트 무인기를 위한 효율적인 안개제거

정수웅,김상권,김중욱,이상근 한국항공우주학회 2013 한국항공우주학회 학술발표회 논문집 Vol.2013 No.4

실외의 환경에서 운용되는 무인항공기(UAV)로부터 촬영된 영상은 안개, 연기, 황사 등과 같은 대기의 입자의 영향으로 인해 쉽게 열화가 발생한다. 본 논문에서는 열화가 발생된 영상의 개선을 위해 안개를 제거하는 방법을 제안한다. 일반적으로 단일영상으로부터 안개를 제거하기 위해선 대기의 빛과 빛의 전달량을 추정해야 한다. 제안하는 방법에서 대기의 빛을 추정하기 위해 Dark Channel Prior 의 방법과 로컬 엔트로피의 방법을 사용하고, 효율적으로 빛의 전달량을 추정하기 위해 로그함수를 사용한다. 또한 스마트 무인기에서 촬영된 항공영상에 적합한 가중치를 적용하여 빠르며 효과적으로 안개를 제거하는 방법을 제안한다. 실험결과에서 제안된 방법의 결과가 기존의 단일영상 안개제거 방법에 비해 효율적이었으며, 항공영상에 사용되기 적합하다는 것을 확인 할 수 있었다. The captured image by UAV(Unmanned Aerial Vehicle) in outdoor environment is degraded by the atmospheric particles such as a fog, smoke, and yellow-dust. In this paper, we propose a defogging algorithm to remove the atmospheric particles. In general, defogging algorithms require the estimation of atmospheric light and transmission. We employ a dark channel prior and the local entropy for atmospheric light estimation. Also, we use a logarithmic function to estimate the transmission. Furthermore, we apply the weighting value for the captured image from UAV effectively. The experimental results demonstrate that the result of the proposed method is more effective and faster than the existing single image defogging algorithm.

비디오 영상에서 SVM과 HMM을 사용한 얼굴영역 추출

정수웅 김천과학대학 2007 김천과학대학 논문집 Vol.33 No.-

In this paper, we present the method of detecting some face regions in video images using Support Vector Machine and Hidden Markov Model. Face images from a camera are mapped into a texture space and then, some texture informations are extracted in face images. Given texture informations, face regions are detected. For extracting some texture informations in face images, 10 Support Vector Machines are used. PCA(principle Component Analysis) is used to reduce dimension, and then a face pattern is obtained. The obtained pattern doesn't represent the entire image texture, but represent the pixel texture information. Accordingly the pattern represents a spacial distribution of each texture information. The detection of face regions is performed using HMM, which shows texture of input image and spatial characteristics of texture. In order to reduce the computational load to the human face detection, the presented method is based on the variation regions application. The method consists of two steps: moving region detection and face region detection. First, the moving region detection, the rough positions of moving objects in image sequence are determined using an adaptive thresholding method that automatically choose the threshold value for detecting the moving regions. Then, we obtained binary motion masks. Second, the face region detection, the pixels in the detected motion masks that have similar texture are extracted. Some candidate face regions are classified into real face regions using HMM. The experiment is showed that the presented method is robust to a motion-blur and saturated image under illumination variation.

Rugby Football 攻擊 Pattern 의 比較硏究

鄭秀雄 嶺南大學校附設 基礎科學硏究所 1987 基礎科學硏究 Vol.7 No.-

The purpose of this paper is to compare the attacking patterns which appeared in the semi-final and final games of the First World Cup Tournament with those which were displayed both in the final game of the 10th Asia Tournament in 1986 and in the final game of 23th Japan Tournament in 1986. The major findings are as follows: (1) As for 'kick off' attacking patterns in the special play, the teams of the World Cup Tournament have showed the 'kick' towards 'touch in-Goal area' most frequently(66.7%). On the contrary, the teams participating in the Asia and Japan Tournament have used the 'kick' towards '10-meter effect area' most frequently(72.2%). (2) As for 'scrum' attacking strategies in 'the broken play', the attacking patterns by means of 'touch kick' have been adopted most frequently(35.5%) in the World Cup Tournament. In the Asia and Japan Tournament, however, the attacks towards the 'open side' by means of 'passing' have employed most frequently(44.8%). In addition, concerning the 'line-out' strategies in broken play, the attacking patterns driving towards the 'ruck and maul side' have been demonstrated with the highest frequency(54,1%), whereas the attacking strategies passing towards 'open side' have showed the highest frequency (31.2%) in the Asia and Japan Tournament. (3) As for attacking patterns in the area of 'ruck nad maul' in 'brea k-down play', the attack towards 'open side' has been highest in frequency(39.1%) in the World Cup Tournament, whereas 'kick and rush' attack has showed the highest frequency(39.0%) in the Asia and Japan Tournament. Besides, as far as attacking patterns of 'catch and pick up' in 'break-down play' are concerned, the attacking strategy of 'touch kick' has appeared most frequently both in the Wored Cup Tournament(51.1%)and Asia and Japan Tournament(67.4%).

유비쿼터스 컴퓨팅을 위한 비전기반의 장소 인식 시스템

정수웅 김천과학대학 2005 김천과학대학 논문집 Vol.31 No.-

Position identification and place recognition is necessary in ubiquitous computing circumstance. In this paper, vision-based place recognition for ubiquitous computing is presented. Video image from a camera is mapped into texture space and texture informations is extracted in image. Given texture informations, place recognition is executed. For extracting texture informations in video image 10 SVM(Support Vector Machine) is used. PCA(principle Component Analysis) is used to reduce dimension, and then place pattern is obtained. An obtained pattern don't represent an entire image texture, but represent a pixel texture information. Accordingly the pattern represent a spacial distribution of each texture information. Place recognition is performed using HMM, which shows texture of input image and spatial characteristics of texture. The experiment is showed that the presented method is robust to a motion-blur and saturated image under illumination variation.

실시간 비디오 감시에서 신경망 기반의 얼굴 검출

정수웅 김천과학대학 2003 김천과학대학 논문집 Vol.29 No.-

본 논문에서는 실시간 비디오 감시 시스템에서 움직이는 사람들의 얼굴을 추출하고 추적하는 방법을 제안한다. 얼굴 추출 과정에서 계산량으로 인한 부하를 줄이기 위해 제안한 방법은 변화 영역 접근 방식에 기반을 두고 있다. 전체적인 얼굴 추출 방법은 두 단계로 구성되어 있다. 첫 번째 단계는 움직이는 영역 추출을 위해 자동으로 임계값을 선택하는 적응 임계치 방법을 사용한다. 그리고, 임계 처리된 영상에서 이진 모션 마스크들을 생성한다. 두 번째 단계에서는 얼굴 영역을 검출한다. 얼굴 영역 추출은 추출된 모션 마스크 내에서 미리 정의된 피부색 모델을 활용하여 유사한 피부색을 가지는 픽셀들이 추출되고 후보 얼굴 영역들이 얻어진다. 후보 얼굴 영역들에 대해 얼굴의 크기와 모양을 학습한 신경망에 적용하여 실제 얼굴을 검출한다. 현재 프레임에서 검출된 얼굴을 바탕으로 연속된 프레임에서 얼굴들이 추적된다. 실험을 통하여 제안한 방법이 잡음이나 배경이 복잡한 영상에서 매우 효과적으로 얼굴을 추출함을 보인다. In this paper, we present a method to track and detect faces of moving humans for real-time video surveillance. In order to reduce the computational load to the human face detection, the presented method is based on the variation regions application. The method consists of two steps: moving region detection and face region detection. First, the moving region detection, the rough positions of moving objects in image sequence are determined using an adaptive thresholding method that automatically choose the threshold value for detecting the moving regions. Then, we obtained binary motion masks. Second, the face region detection, the pixels in the detected motion masks that have similar skin-color are extracted using the predefined skin-color model, which is a stochastic model to characterize skin-colors of human faces, then we obtain the binary candidate face region masks. The candidate face regions are classified into the real face region using neural network, which is trained size and shape of face. Based on the detected face in current frame, faces are tracked in consecutive frames. Our experimental results show that the presented method is robust under complex background and some noise.

On-Line 한글 인식을 위한 퍼지 이론의 이용에 관한 연구

정수웅 김천과학대학 1993 김천과학대학 논문집 Vol.19 No.-

In this paper, Using a Fuzzy theory an On-Line phoneme recognition procedure of Korean characters is studied. The recognition method is performed by reading the pixel coordinates, finding the frature points, recognizing the stroke, recognizing the phoneme in order. In the course of recognizing the stroke, First the frature points of an inputed stroke is characterized by starting point, end point, and bending point. Second, We investigate that The connection style between the former stroke and the latter stroke from the set of input strokes is related. And The relative direction between the former stroke and the latter stroke is considered. Finally, According to the above three form(stroke code, connection style, relative direction) We recognize a written korean character using finite automata.

영숫자 그림이 혼용된 한글문서에서 문자 분리에 관한 연구

정수웅,이갑래 김천과학대학 1998 김천과학대학 논문집 Vol.24 No.-

In this paper, we propose a new method for segmenting characters in hangul document mixed with alphanumeric characters and picture. Since hangul has structural characteristics different from those of alphanumeric characters, structural characteristics of hangul characters are also different from those of alphanumeric ones. If hangul and alphanumeric characters are both written in a document, it is difficult to know whether the touching characters are hangul or not. The proposed segmentation method uses an MLP to generate candidate cutting points. The MLP-based segment lea군 cutting points from training samples which are composed of features extracted from touching character images and correct cutting point of those images. It generates five candidate cutting points per a touching character image and each candidate has a value that is regarded as cutting possibility at that position.