Technologies And Realization Of A Practical Face Recognition System

With the development of society and technology, the demand of effective automaticrecognition technology is increasingly urgent. Compared with traditional lock and pass-word, facial features are more convenient, reliable and safe. Compared with other humanbiological features, such as finger and iris, facial features are more friendly and easy to trace.There is no doubt that while this technology become more and more mature, it will greatlychange the life pattern of human beings.Face recognition technology has been developed for many years. Up to now, these aremany recognition algorithms. But a practical face recognition system still face with manydisturbances which have not been solved. Technologies such as illumination normalization,face feature extraction and selection play important roles in improving the stability of thesystem. In my thesis, I focus on these two key technologies, propose new recognition meth-ods and improve some existing ones. Based on these algorithms, I will also introduce apractical system that has been accomplished by my team. The main work of this thesis canbe described as follows:·Logarithmic Total Variation (LTV) model can decompose the face image by featurescale, and extract light invariant face features. But its results, especially on imageshad high gray level, have low contrast ratio and the gray level are not normalizedamong these results. In my thesis, an improved method will be presented to solve theproblems in LTV model, using nonlinear transform to improve the contrast ratio ofthe result images, get better face features and normalize the whole gray level. Theexperiment results on Yale Face Database B and CMU PIE Face Database show thevalidity of the proposed method. It improves the recognition performance of LTVmodel under uneven illumination.·The accuracy of face recognition not only depends on the classifier, what feature willbe used to describe different faces is more important. So my thesis studies on feature extraction and selection. Propose a face geometric extraction method based on ELTV.This method can automatically extract face geometric information and use Hausdorffbased method to fuse geometric and texture information. Also conclude the featureselection methods based on AdaBoost, especially Gabor and LBP feature selection.These two methods both show good results in recognition.·Based on the research of algorithm, this thesis will also give a full solution of a facerecognition application. The system use face recognition as the core, can be used tomanage a smart district and provide many functions such as access control, attendancecheck, exit and entry records. It can also be used in many occasions like company,school and examination management. It provides a good platform to test the effective-ness and stability of the algorithm and also lies a solid foundation for the practical useof face recognition technology.