Robust Linear Represenation Models And Algorithm Design For Face Recognition

As one of the most promising branches of pattern recognition, the facerecognition technology has attracted tremendous attention from major researchgroups and also becomes the most important part of biometrics. Face recognition hasachieved enormous development in the past few decades, and the linearrepresentation based classification method?LRBCM? is one of the most strikingface recognition algorithms, because it has extraordinary priorities in solving facerecognition problem. However, LRBCM still cannot perfectly solve the facerecognition problem, mainly because its recognition accuracy is always influencedby external environments, such as variations in poses and expressions, differentillumination conditions, occlusion and disguise problems. Furthermore, these issuesstill need to be urgently resolved for face recognition. This dissertation proposedtwo robust linear representation based classification frameworks for facerecognition.This dissertation first proposed a noise modeling and linear representation basedclassification method. This method first models the representation noise, and thensimultaneously diminishes the representation noise in each iteration and achieves abetter linear representation solut ion, and finally exploits the multiple classifierfusion strategy to perform more accurate classification. LRBCM generally consistsof two categories, i.e. the global LRBCM and the local LRBCM. In order to obtainthe bilateral dominant characteristics of the globality and locality of LRBCM, weproposed to integrate global LRBCM and local LRBCM together, i.e. the secondmethod in this dissertation, to perform robust face recognition.Experiments on several public face databases, such as AR, FERET, ORL, GT,CMU PIE and corrupted face databases, have been conducted to test the frameworksproposed in this dissertation. Meanwhile, some recently proposed powerfulLRBCMs are utilized for comparisons. The experimental results demonstrated thatthe proposed frameworks have highly computational efficiency and also candramatically improve the robustness and classification accuracies.