Face Recognition Based On Multi-Camera System

Recently, face recognition has become a hot research topic in computer vision andpattern recognition. As one of the most friendly biometric authentication methods, facerecognition has wide applications in many areas, such as public safety, military security,national security and so on. However, there still exits some great challenges in facerecognition applications. On one hand, since the result of face recognition is impacted bymany factors, such as variation of ambient light, the postures of heads, and the changes offacial makeup, how to make the face recognition system robust against these factor is animportant problem in theoretical research and practical applications. On the other hand, toreduce the cost and complexity of face recognition system is also critical for the developmentand application of face recognition technology.Existing face recognition approaches can be generally divided into two categories, i.e.,three-dimensional (3D) face recognition and two-dimensional (2D) face recognition.However,2D face recognition approaches are generally affected by light, postures and otherfactors.3D faces data collection equipment is relatively expensive, and the composition of3Dface recognition system is quite complex. Therefore, in this dissertation, a face recognitionapproach that combines the3D rotation model and the improved SIFT (Scale InvariantFeature Transform) algorithm is proposed.The main work of this dissertation can be summarized as follows:1. A comprehensive survey of face recognition is carried out in this dissertation. Severalmajor methods of face recognition are summarized, including face recognition based onelastic graph matching, subspace-based face recognition, face recognition based on neuralnetworks, face Recognition Based on Hidden Markov Model and three-dimensional facerecognition. Meanwhile, problems in the3D and2D face recognition are discussed.2. A novel face detection method is proposed by combining Adaboost algorithm andskin color model. This approach is more robust to the variation of background and colors, andits detection speed is fast. We test the proposed approach on both CMU PIE database offeredby Carnegie Mellon University and an in-house database collected from our multi-camerasystem. Experimental results show that our proposed method could achieve high recognitionaccuracy.3. Principal component analysis (PCA) and Linear Discriminant Analysis(LDA) facerecognition approaches are described in detail in this dissertation. We also present acombination of PCA and LDA face recognition methods, which reduces the feature dimension and effectively solves the two-class classification problem. Moreover, this approach is simpleand fast.4. In this dissertation, the SIFT feature extraction algorithm is described and animprovement of SIFT feature point matching is proposed. According to the variation ofpostures, some information of faces are lost in the processing, hence, the SIFT algorithmcouldn’t achieve enough feature points for matching. In order to make the SIFT algorithmrobust against face postures, we present a novel approach that combines3D rotation modeland the conventional SIFT algorithm.5. We have constructed a multi-camera face recognition system which is robust to thevariation of face postures. Meanwhile, a multi-camera systems face database has also beenestablished. The database includes data from41volunteers, including three groups of gesturesfor each person, i.e., looking up, looking horizontally and looking down, and each groupcontains left view, front view and right view, a total of9photos. It also has various ofexpressions for each person.