Research On The Image-based Face Recognition

Biometric recognition is a new kind of personal identification technology and facerecognition is the most popular one of biometric recognition approaches. Because of facerecognition's high acceptability and low hardware requirements, it has been researched deeply.In recent decades, face recognition has achieved extensive application. Face recognition inreal life has wide application prospects. It is of great significance in authentication,information security, driving safety, crime investigation, intelligent man-machine interface,mental illness diagnosis, etc. It can be said face recognition has become an important part inImage Engineering and Pattern Recognition fields. Many excellent recognition algorithmshave improved the recognition accuracy into the new levels from time to time.However, in fact, there are still many unsolved problems in this field. Face recognitiontechnology has some existing shortcomings. For example, recognition accuracy rate is notideal, the reliability is poor, and most of the current face recognition methods have a certaindistance from the commercial application standards. Thus, study the more effective facerecognition methods, improve their recognition accuracy, enhance their robustness and theenvironment adaptability, are still has huge theoretical significance and application value.This paper focuses on some key issues in the image-based face recognition method, andgive out some deeply study and explore results, such as face alignment, effective local andglobal face feature extraction, multi-subclass distribution classification, low-separableproblem and small sample size problem. In this paper, the problems mentioned above in theface recognition have been studied in theoretical and experimental demonstration, andachieved some progress. Specifically including:Face alignment. According to the eye region's geometric structure and gray distribution, anew integral projection function CNMDI-IPF for eye location is proposed. The first, it usesradial symmetry transformation (RST) extract the facial regions with high radial symmetricnature, then it combines with gray integral projection and facial organs' distribution todetermine the candidate eye regions. The second, it computes average eye template, andemploys image Euclidean distance as distance measure to complete the template matching ofcandidate eye regions. This process could get the real eye region. The third, in the pupil location stage, we define a new integral projection peak evaluation function to quantify thepeak (valley) evaluation value. The accurate pupil position can be obtained according to thesorted evaluation value. For the proposed method takes into account the texture and structuralcharacteristics, so it can precisely locate the pupil. This method is able to eliminate the errorcaused by face misalignment in the identification process, so it is effective to improve the facerecognition rate.In local feature extraction of face image research, an improved weighted LBP codingpattern RDW-LBP is proposed. RDW-LBP is designed according to the impact of statistics ofpixels in local neighborhood with different positions. Moreover, RDW-LBP is used indifferent scales and directions discrete wavelet decomposition coefficients to extract thefeature map. This will not only take into account the influence caused by different sub-regionsof the human face, but also consider the texture details aspect in face recognition. In addition,for the local ternary pattern feature usually has large dimensions, a low-dimensionalhistogram sequence to describe the face appearance method, called LTP sub-pattern (LTP-SP)is proposed. LTP-SP can be considered as the Fisherfaces of face image sample's LTP map.This algorithm can reduce the facial feature dimensionality as well as enhance the sampleseparability.Aiming to solve the small sample size problem in face recognition, a new face recognitionmethod using improved complete fuzzy membership and downsampling bidirectional2DLDAensemble (ICFM-DBi2DLDAE) is proposed. In the proposed method, we give out threecontributions as follows. First, we use linear regression to simplify the complete fuzzymembership definition. The simplified definition is able to achieve a comparable performancewith the original definition, and avoid cross validation. Second, we give the theoretical proofand experimental support to further discuss the essence of2DLDA and A2DLDA. Theinformation extracted by these two kinds of matrix-based LDA methods have consistency andcomplementarity. Third, we employ image downsampling to enhance the independent andidentically-distributed situation of samples. Column-based and row-based classifierensembles are used to stabilize the recognition performance. This stage has not only avoidedthe limitation of single-directional information, but also resolved the SSS problem.Experiments on ORL, AR and PolyU-NIR databases show that our method works well underdifferent conditions, and it is superior to other discriminant analysis methods.Multi-subclass distribution classification problem is usually occurred in real life. In thispaper, two aspects of the multi-subclass discriminant analysis algorithm is discussed. The firstone, we analyze the problem that the criterion function of multi-subclass discriminant analysisalgorithm and the Fisher criterion is not equivalent. This will lead to difficult to achieve thetheoretical minimum Bayesian error of discriminant analysis algorithm. For solving the problem, a revised criterion function is given out. The second one, in order to solve theclassification problem that image samples are multi-subclass distributed and they arenon-linearly separable, a Kernel2-Dimensional Subclass Discriminant Analysis (Kernel2DSDA, K2DSDA) algorithm is proposed. In this section, we give out the theoretical supportof that K2DSDA is equivalent to column/row-2DSDA based on kernel image samples.Secondly, kernel between-subclass scatter matrix and kernel covariance matrix are computedvia the approximate definition of kernel samples. So the computational complexity of theproposed algorithm is reduced enormously. At last, the proposed method is tested onbenchmark face databases and the experimental results show that our method is superior toother state-of-the-art discriminant analysis algorithms, which confirms the effectiveness ofK2DSDA.For the "Suboptimality" problem often appears in the discriminant analysis algorithms, animproved weighted Mahalanobis distance definition is proposed. The new Mahalanobisdistance definition considers both of weight influences caused by sample scatter degree andsample size, so the classification projection axes more focus on large-scale, highly discreteand difficult classification categories. The weighted Mahalanobis distance is combined withsubclass discriminant analysis algorithm framework. This method is able to solve theclassification problem that the samples are subclass-distribution and there are largedifferences between these subclasses.Face recognition is a challenging research topic. There are still many difficult points in reallife face recognition. The paper's work is meaningful for single mode face recognition usinglow-cost equipment.