Dissertations / Theses on the topic 'Multiple face detection'

Afin d’améliorer les performances d’un problème de classification, une piste de recherche consiste à utiliser plusieurs classifieurs et à fusionner leurs sorties. Pour ce faire, certaines approches utilisent une règle de fusion. Cela nécessite que les sorties soient d’abord rendues comparables, ce qui est généralement effectué en utilisant une calibration probabiliste de chaque classifieur. La fusion peut également être réalisée en concaténant les sorties et en appliquant à ce vecteur une calibration probabiliste conjointe. Récemment, des extensions des calibrations d’un classifieur individuel ont été proposées en utilisant la théorie de l’évidence, afin de mieux représenter les incertitudes. Premièrement, cette idée est adaptée aux techniques de calibrations probabilistes conjointes, conduisant à des versions évidentielles. Cette approche est comparée à celles mentionnées ci-dessus sur des jeux de données de classification classiques. Dans la seconde partie, le problème d’anonymisation de visages sur des images, auquel SNCF doit répondre, est considéré. Une méthode consiste à utiliser plusieurs détecteurs de visages, qui retournent des boites et des scores de confiance associés, et à combiner ces sorties avec une étape d’association et de calibration évidentielle. Il est montré que le raisonnement au niveau pixel est plus intéressant que celui au niveau boite et que, parmi les approches de fusion abordées dans la première partie, la calibration conjointe évidentielle donne les meilleurs résultats. Enfin, le cas des images provenant de vidéos est considéré. Pour tirer parti de l’information contenue dans les vidéos, un algorithme de suivi classique est ajouté au système<br>In order to improve overall performance of a classification problem, a path of research consists in using several classifiers and to fuse their outputs. To perform this fusion, some approaches merge the outputs using a fusion rule. This requires that the outputs be made comparable beforehand, which is usually done using a probabilistic calibration of each classifier. The fusion can also be performed by concatenating the classifier outputs into a vector, and applying a joint probabilistic calibration to it. Recently, extensions of probabilistic calibrations of an individual classifier have been proposed using evidence theory, in order to better represent the uncertainties inherent to the calibration process. In the first part of this thesis, this latter idea is adapted to joint probabilistic calibration techniques, leading to evidential versions. This approach is then compared to the aforementioned ones on classical classification datasets. In the second part, the challenging problem of blurring faces on images, which SNCF needs to address, is tackled. A state-of-the-art method for this problem is to use several face detectors, which return boxes with associated confidence scores, and to combine their outputs using an association step and an evidential calibration. In this report, it is shown that reasoning at the pixel level is more interesting than reasoning at the box-level, and that among the fusion approaches discussed in the first part, the evidential joint calibration yields the best results. Finally, the case of images coming from videos is considered. To leverage the information contained in videos, a classical tracking algorithm is added to the blurring system

Thesis (Ph.D)--Electrical and Computer Engineering, Georgia Institute of Technology, 2009.<br>Committee Chair: Rusell M. Mersereau; Committee Member: Biing Hwang (Fred) Juang; Committee Member: Christopher E. Heil; Committee Member: Georgia Vachtsevanos; Committee Member: James H. McClellan. Part of the SMARTech Electronic Thesis and Dissertation Collection.

碩士<br>國立交通大學<br>資訊學院資訊科技產業專班<br>98<br>Face detection is an important research topic in recent years. Automatic face detection is applied to many applications (e.g., Surveillance System, Security System, Video Coding, Video ROI (region of interest) rate control, Video Conference, information retrieval in video or images). We hope to find a good method of face detection. In this thesis, we proposed a face detection system based on multiple criteria. We use many points of view to consider how to do face detection in color images. We use not only local rectangle features but also global information about faces to construct face classifiers. We also use the human skin color features to decrease false alarms in color images. We introduce our proposed method in detail in this thesis. And we compare our proposed method with a few existent methods. We can know that our proposed method shows higher correct detection rate and lower false alarm rate than other existent methods from comparing table. In conclusion, we proposed a face detection system with higher correct rate and lower false alarm rate in this thesis.

碩士<br>義守大學<br>資訊管理學系碩士班<br>98<br>Today, images have been ubiquitous, such as the images in television, movies, photos and mobile phones, etc. In a variety of data type images occupy the most storage space than other. When the image data is stored more and more as time, it will take a lot of time to find out an image you want from an image database. Therefore, how to quickly recognize the content of images becomes very important. For the security department, how to utilize human face in image to quickly find out the corresponding data is very important. It is a difficult matter that how to let the computer know whether there are human faces in an image. Because the contents of the image are very complexity, so for recognizing human face in an image, the computers have to do some processes including searching, locating, analyzing, and recognizing processes. The first process of human face recognition is face detection. Due the accuracy of face detection and location can greatly affect the result of face recognition, therefore, how to develop an effective method of face detection is very important. Because multiple-face detection is more difficult than single-face detection, therefore, the goal of this study is how to correctly detect multiple faces in an image. This study utilizes following methods: color segmentation, light compensation, hole technology and the erosion and dilation method of the morphology, to clear non-skin area in an image. This research proposes a scheme of skin reanalysis and uses the statistical features of holes to clear the skin not belonging to human face. After that, using Soble operator and lips-frequency color analysis complete multiple-face detection.

碩士<br>元智大學<br>資訊管理學系<br>98<br>Human face detection and human face recognition are always popular topics in pattern recognition, they have develop for a long time, and its technique has become more and more robust. At the past, some technique could be restricted by hardware or other experimental settings, so they should made some kind of trade off between detection rate, recognition rate and computing time. Now days, with the development of technology and other equipment some time consuming methods become available. We present a human face detection method which is base on haar-like features and integral image and a human face recognition method which divide human face into different feature area, then use machine learning to divide the feature groups and classify them by these feature information.

碩士<br>國立臺灣師範大學<br>資訊工程研究所<br>98<br>Face detection and human detection are important in all surveillance method applications. In classroom, we can use detection to assist us to observe student activities. Their response will give some suggestions to teacher, and teacher can improve the teaching. Furthermore, it can extend automatically real-time roll call system to help teacher. We propose a new detection method in classroom. Our method employ a combination of AdaBoost classify faces, applied filter and HOG find trustworthy human face. Bubble-Developing Mechanism (BDM) is a similar object tracking method. It’s an easy way to solve the continuous problem in video sequence or live video. Bubble means individual face results in each of frame and they will have weights just like age. Growth over time, bubbles grow old or die. Because BDM have characteristics of time and continuous, it can enhance the performance of our method. In experiment results, improve AdaBoost and applied filters have a better frame rate than original AdaBoost for real-time face detection. BDM can achieve detection rate from 72% to 94% in single person detection and have average 85% detection rate in multiple people environment.

碩士<br>國立臺灣大學<br>資訊管理學研究所<br>93<br>The challenges for face detection from images come from the variation of poses, facial expressions, occlusions, lighting conditions, and so on. We propose a method for multiple-pose face detection from still images. Our proposed method consists of three phases. First, skin pixels are extracted using a skin color model. Connected component analysis is performed to find the skin regions. Second, before extracting the feature vector of a skin region, we apply edge detection to the region. Our feature vector consists of two parts. The first part is obtained by dividing the edge image into 3*4 grids and calculating the number of horizontal edges and the number of vertical edges in each grid. The other part is obtained by computing the summary of color correlogram of the edge image. Third, with a set of training images, the fuzzy c-means (FCM) clustering algorithm is used to build face models. If the Euclidian distance between a feature vector and a face model does not exceed a predefined threshold, the region will be classified to a face. The experimental results show that our method can deal with the variation in poses, rotations, scales, and so on.

With the generalized use of systems for digital contents dissemination arises the opportunity for implementing solutions capable of evaluating audience reaction. This dissertation reflects the implementation of one of those solutions. To this end, the development involved adapting a previously functional digital signage system. In this sense, digital cameras were paired to the content display terminals in order to capture information from the area in front of them. Using computer vision technologies, the terminals detect, in real time, people who appear in the cameras’ field of view, and this information is communicated to a server for data extraction. On the server, methods are used to perform face and emotion recognition, and also to extract data indicating the position of the head, which allows the calculation of an attention coefficient. The data is stored in a relational database, and can be consulted through a web platform, where they are presented associated with the contents corresponding to the moment of their capture and extraction. This solution thus allows the evaluation of the impact of the digital contents presented by the system.<br>Com a utilização generalizada de sistemas de disseminação de conteúdos digitais, surge a oportunidade de implementar soluções capazes de avaliar a reacção do público. Esta dissertação reflete a implementação de uma dessas soluções. Para isso, o desenvolvimento passou pela adaptação de um sistema de sinalização digital previamente funcional. Neste sentido, aos terminais de exposição de conteúdos, foram emparelhadas câmaras digitais de modo a permitir a captação de informação da área à frente destes. Com recurso a tecnologias de visão de computador, os terminais fazem, em tempo real, deteção de pessoas que apareçam no campo de visão das câmaras, sendo esta informação comunicada a um servidor para extração de dados. No servidor, são utilizados métodos para realização de reconhecimento de faces e emoções, e também é feita extração de dados indicadores da posição da cabeça, o que permite o cálculo de um coeficiente de atenção. Os dados são guardados numa base de dados relacional e podem ser consultados através de uma plataforma web, onde são apresentados associados aos contéudos correspondentes ao momento de captação e extração destes. Esta solução, permite, assim, a avaliação do impacto dos conteúdos digitais apresentados pelo sistema.<br>Mestrado em Engenharia de Computadores e Telemática

碩士<br>國立交通大學<br>應用數學系數學建模與科學計算碩士班<br>102<br>An image for people, usually they only interest some parts of it, these parts called "target" or "front scene", and the other parts called "background". Dividing an image into targets and back scenes is the basic work in image processing, this work has primarily goal which is to recognize the edges and boundaries of a target. For human's visual system, that can easily identify a target whose edges and separate target from background, which is an easy task to accomplish. However, in computer’s visual system, how to imitate human's visual ability, is a study-worth topic. In this paper, we’ll using Robert, Prewitt, Sobel and Laplacian, four methods of edge detection to extract edge from human face, and combine Mean, P-Tile, Two-Peaks and Optimal Threshold Iterated, four image segmentation threshold algorithm to increase the results of edge detection, make the effort better. In furthermore, to compare with these different methods’ pros and cons.

碩士<br>國立交通大學<br>電子工程學系 電子研究所<br>103<br>This thesis proposed a multiple symptoms image detection system, where the functions of the system include ulcer detection and face skin detection such as spot detection and redness detection. A common method for ulcer detection used machine learning. However, the machine learning method would require much time for training and collecting positive and negative image samples. Instead of using machine learning methods, we build up a multiple symptoms analysis system to support doctors to quickly review all endoscopy images to find out the images containing suspicious symptoms, which can reduce doctor’s loading and save much time. Moreover, we also propose some face disease features to build up face skin detection, which provides the customers to examine and care their face condition in any time. There are some common problems in the existing researches. For example, many algorithms can detect ulcer and face diseases in balance illumination environment because they cannot handle arbitrary reflection and unbalance light. The proposed system can overcome the above mentioned problems and execute correctly. The proposed system is developed and implemented on PCs with the performance about D1 video (720x480) at 15 frames per second.

碩士<br>淡江大學<br>機械與機電工程學系碩士班<br>98<br>The object of this thesis is to design the assistive communication device for visually impaired people with multiple disabilities using face detection and tracking systems, to solve the problems that visually impaired people with multiple disabilities can’t communicate and express to outside world with words. In the study, the images of head shaking are captured into a notebook through a CCD camera with IEEE-1394 interface or a USB webcam. The directions of head shaking are identified by face detection and tracking systems. Then the voice communication with the outside world can be carried out using text-to-speech editor. 　　At first, the face detection and tracking is performed by the CAMShift algorithm in the study. The rotation angle of the face is computed through the moving position of face center to identify the directions of head shaking, and then the communication operations of Pinyin, associating Chinese character, and debugging error are processed. Finally, the voice communication and text output are achieved at the same time by the integrated text-to-speech editor. The assistive communication device is accomplished for visually impaired people with multiple disabilities using face detection and tracking systems. The results of this thesis can achieve that visually impaired people with multiple disabilities do not need to wear or touch any auxiliary device to operate properly. In addition to the increased use of convenience, it can also improve the visually impaired people with multiple disabilities to communicate with the outside world and to create the better learning and living environment.

碩士<br>國立成功大學<br>工程科學系碩博士班<br>93<br>In this thesis, an efficient face identification system is proposed. The system can be applied to a security monitoring, doorway intercom and interactivity service, etc. The identification system mainly consists of two sub-systems: one is face detection and another is face recognition. The face detection sub-system detects face search region using skin color information. We detect the possible face images using an ellipse template algorithm, and finally the human face using principal component Analysis (PCA). For the face recognition sub-system, we can extract the low-dimensional discriminative feature parameter for human faces using the discrete wavelet transform (DWT) and Linear Discriminant Analysis (LDA). Finally, we employ the nearest feature line (NFL) to determine the most likely person. We can construct a robust and high-accuracy recognition system. The experimental results in face detection part show that the successful rate is 98.4%. For the face recognition part, the recognition rate for a single image reaches 94%. Finally, the computation time of the entire face recognition system is 0.26 seconds on the average, using a Pentium M 1.5G personal computer.

碩士<br>國立雲林科技大學<br>電子工程系<br>105<br>The broad applications of face detection in numerous practical problems have brought prosperous research effort in recent years. Although much progress has been witnessed, the complete face detection solution is still not available. The obstacles lying in real situations are typically lighting, camera angles, face poses, facial expression, occlusion, multiple faces, skin complexity, background clutter, etc. This work adopts the skin color-based approach, due to skin color's invariance to scale, pose, camera angle, and ease of implementation. Aiming at reducing the false alarm rate in detecting images containing multiple faces, our scheme incorporates several ingredients: (1) a GMM model for the evaluation of the probability of a skin pixel, (2) a seam carving mechanism for the separation of multiple faces, and (3) MouthMap and EyeMap for the detection of mouth and eyes in a face to eliminate non-face color regions. The proposed method is tested on well-known face databases, and the result indicates good detection performance, even when the input image consists of a number of partially overlapped faces.

碩士<br>國立東華大學<br>資訊工程學系<br>90<br>In this paper , we present a rule-based methods for real-time multiple faces detection. First, the system will find face candidate regions with skin color pixels in images or video frames by using checking rules built on a quadratic polynomial model. After extracting the face candidate regions, the lip and eyes are located by using some geometrical rules. Finally, the precise location of the face region are determined based on the extracted lip and eyes. Due to the simplicity of our algorithm, the whole face detection process can detect faces within various complicated environment conditions at a real-time speed.

碩士<br>崑山科技大學<br>電機工程研究所<br>94<br>This thesis utilizes Pan-Tile-Zoom camera (PTZC) to develop a multiple faces sequential detecting and tracking system. In the face detection part, first we extract skin region from image by color segmentation, then utilize the methods of ellipse matching and connected components labeling to decide the candidate face regions. After the preliminary face positioning, we use some geometry rules to search face features (eyes, lips) in order to find out the correct position of the face candidate. In the face tracking part, we use PTZC to track the moving face candidate, keep it in the picture center and room in PTZC to enlarge the face candidate. Here we also use the image projection method to remove those skin color pixels which are not belonged to the face’s elements from the image to get more robust dynamic face tracking results. 　　For multiple face sequential tracking, the system will be interested in second face candidate when it appears. We track the new coming face candidate and let it become our new target. The same detecting and tracking methods will be used on this new face candidate. Moreover, the system will control PTZC to focus on this target. The same rules will apply on next coming new face candidates. When the face candidate disappears from the vision area of PTZC the system will start to search and track the other face candidate in its vision area continuously. After experiment with the real-time images of multiple faces, the result shows that the system can reach multiple faces detection and tracking, and proves the feasibility of proposed method in this thesis.