Face Recognition Using Supervised Independent Component Analysis

In the daily life, face contains and conveys mounts of information; therefore it holds an important situation in our daily communications. Due to the unique advantage and research value, face recognition is always an emphasis and hotspot in the field of pattern recognition. Also, it has been used in many occasions.After searching, reading and analyzing lots of papers about face's research, this thesis puts the emphasis on the subspace methods in face recognition. Furthermore, it pays the major attention to researches on face recognition based on the independent component analysis (ICA). Theoretically, comparing to the method of principal component analysis, which has been used for many years, ICA can expose the essence behind the face images. So, ICA has more advantages on feature extraction. The main work of this thesis includes the following aspects:1. The chapter one briefly introduces the basic methods in face recognition, and chapter two presents some basic concepts and algorithms in ICA. Following these basic concepts, we also present the history, hotpots and extensions of face recognition and ICA, respectively.2. After the analyzing of traditional ICA method, the thesis uncovers its disadvantage in classification. Taking the full use of the class information, we present a supervised version of ICA (SICA). Here, we do not only rest the class information on the selection of independent component. Furthermore, it mixes the information in the iteration of learning. By introducing class scatter degree within class and using it as a restriction condition, we derive a new learning rule. Experiments on three standard face databases prove that SICA is feasible and efficient.3. At last, the thesis discusses the combination of the features, which are extracted by ICA, and sorts of classifiers. Comparing the most prevalent classifiers nowadays, i.e. the nearest neighbourhood classifier and support vectormachine (SVM), the thesis points out that the features extracted by ICA are more discriminate than those by PCA. Also, experiments on ORL face database verify the conclusion that the combination of ICA and SVM gets the highest rate of recognition.