Real-time face localization and recognition system by using video sequences

In recent years face detection and face recognition has received substantial attention from bothresearch communities and the market, but still remained very challenging in real-timeapplications. The machine learning and computer graphics communities are also increasinglyinvolved in face recognition. This common interest among researchers working in diverse fieldsis motivated by remarkable ability to recognize people and the fact that human activity is aprimary concern both in everyday life and in cyberspace.Although some techniques of face detection and recognition are able to produce considerablygood results in some particular environmental conditions, even minor changes in theseenvironmental conditions might have broken effects on these techniques. In the context of thisthesis, necessary methodologies and approaches were studied to design a real-time face detectionand recognition system by using video sequences. Moreover, a robust and practical softwareapplication has been developed to solve these problems.In order to solve face detection problem, about half million skin color pixels were collected fromthe face images taken under different lighting conditions and pose orientations. Thedifferentiation of skin and non-skin pixels on a human image is a binary classification problem.Therefore, the YCbCr color space representation of these collected pixels and the rest of the colorspace has been distinguished by using support vector machine algorithm which has proved itsability and strength in binary classification problems. By using these classified pixels, a novelapproach to detect skin colors on the video frames has been introduced. Then, a statistical facelocalization method has been utilized to detect faces on video frames. This method basicallycalculates the center coordinates, width and height of face areas by using histograms createdalong the x and y-axis on the video frames. As a last step in order the increase the performance offace recognition system, the outputs of face detection system are converted into a standardizedcanonical face model. The success ratio of the face detection system is 90% whereas its failurerate is 10%.The face recognition system derives the feature set of human faces by using the projections ofeigenface space which is a principal component analysis method. The principal componentanalysis method considers the face recognition problem as a 2-D pattern recognition problem. Itfirst creates the set of eigenvectors and eigenfaces of face space which most efficiently extractsthe differences between different face frames. Then the real faces are represented as thecombinations of these eigenfaces and the feature sets of the faces are derived. As a last step, thesimilarities between the feature sets of training and test faces are calculated by using the supportvector machine algorithm. Each subject within the face recognition system is represented via atleast five faces and this approach obviously increases the performance level of face recognitionsystem. In order to get most consistent recognition results, sliding window approach has beenintroduced by using the sequence of video frames The success ratio of the face detection systemis 91% whereas its failure rate is 5%.Keywords: Face recognition, face detection, principal component analysis, support vectormachine, video-based image processing, real-time applications.JURY:1. Assist. Prof. M. Elif KARSLIGİL (Supervisor) Date: 05.06.20062. Prof. Dr. A. Çoşkun SÖNMEZ Page: 903. Assoc. Prof. Selim AKYOKUŞ