Research On Local Features Extraction Method For Face Recognition

Face recognition technology is regarded as one kind of widely accepted biometriciden-tification technology, and it has been widely applied in finance, security, access control andother aspects due to its low cost, simple, friendly interface and a series of other advan-tages.However, there are still many key technology need to be further improved and perfecteddue to the human face images are easily affected by external and internal factors.Face recognition including features extraction and features matching two stages, inwhich features extraction stage is undoubtedly the key stage throughout the recognition proc-ess. Features extraction methods can be divided into two types, one is based on local informa-tion extraction method, and the other is based on global information extraction method. Localfeatures has been widespread concerned in face recognition because of its robustness to thechanges of illumination, expression, and partial occlusion. As a typical local features, Gaborfeatures has been widely used for its robustness to the external environment in face recogni-tion.However, most of the current methods based on Gabor filters presence the problem oflarge computation complexity. In addition, local monogenic information not only includesmore identifying information, but also has the advantages of low computational complexity,and shows a better face recognition application prospects.In this paper, the local features ex-traction methods based on Gabor filtering and Monogenic filtering is studied for the purposeof extracting the most efficient characteristic.we aim at how to extract lower dimension Gaborfeatures from the Gabor information and how to explore the potential of monogenic informa-tion in face recognition to take advantage of the multi-mode monogenic information aftermonogenic filter filtered. Meanwhile,this paper studies how to applied Linear DiscriminantAnalysis to fuse different kinds of feature to enhance the performance of the algo-rithm.Besides,a method for face recognition under complex illumination conditions is alsoproposed.The article carries out the research work of the following aspects:1．In order to reduce the dimensionality of the extracted Gabor features to reduce thecomplexity of the algorithm, a novel face recognition method based on Gabor DirectionalPattern (GDP) is proposed which encoding all the8Gabor magnitude maps with the samescale.This method can reduce the dimension of the extracted features effectively with a highrecognition rate.2．On the basis of the study on the real and imaginary parts of Gabor information, thereal part and the imaginary part characteristic feature are extracted by apling the improvedGDP operator to the real and imaginary part of the Gabor filtered map,and the two kind offeatures are further fused by Linear Discriminant Analysis(LDA) to further improve the rec- ognition performance of the characteristics3．To take full advantage of the multi-mode monogenic information, a new methodnamed Monogenic Local Quantization Patterns (M-LQP) which fused the monogenic magni-tude, direction and phase information on the sample level is proposed base on the study of lo-cal quantization pattern.Because of the three kind of information were fused in M-LQP fea-tures at the level of integration, this method achieved perfect performance with very low fea-ture dimension.4．To compensate for the disadvantages of the Monogenic Binary Pattern (MBP) methodwhich not take full use of the monogenic direction information, a new operator named En-hanced Patterns of Monogenic Orientation Difference (EPMOD) is proposed to get EPMODfeature of the Monogenic orientation information, and then the EPMOD features are fusedwith the MBP feature by Block Based Linear Discriminant Amalysis (BLDA), and thismethod effectively improves the recognition rate.5．In order to improve the performance of face recognition under non-uniform illumina-tions conditions，a face recognition method based on Patterns of Monogenic Oriented Mag-nitudes(PMOM) oprator is proposed.This operator decomposes the monogenic orientation andmagnitude map into several PMOM maps by accumulating local energy along several orienta-tions and this method has improves the performance significantly for the image face with il-lumination variations.In summary, this article presents several kinds of local features extraction methods basedon Gabor and Monogenic information.A number of experiments are conducted over ORL,CAS-PEAL and Yale-B face database, the results show that the several feature extractionmethods proposed based on Gabor and Monogenic information has better or comparable per-formance than traditional local feature methods but with significantly lower time and spacecomplexity.