Research On Feature Extraction Algorithms With Locality Preserving Projection Of Face Recognition

Face recognition, which is an efficient identity recognition, has been attractingworldwide research interests of the experts and scholars. As key component of facerecognition technology, the feature extraction has been become an important researchtopic. To obtain effective features for distinguishment, many feature extractionalgorithms are proposed. In these proposed algorithms, the manifold learning algorithmhas become the research hotspot in the field of face recognition in the21st century dueto its theory stylish and unique perspective. Based on manifold learning, the localitypreserving projections (LPP) grows the typical representatives of feature extraction, andthen develops two dimensional locality preserving projection (2DLPP).In the existing LPP and its improved methods, the variance contribution is regardedas the evaluation criteria of feature extraction quality. These traditional local linearmanifold feature extraction algorithms cannot guarantee the classification performanceafter dimension reduction. Thus, a locality preserving projections feature extractionalgorithm based on Shannon Entropy, which is abbreviated S-LPP, is proposed in thispaper. For the proposed algorithm, the classification uncertainty of the feature extractionis described by the Shannon Entropy that is also regard as the evaluation criteria offeature extraction. In the proposed method, we firstly exploit Shannon entropy to carrythe distinguishment information of classification. Then, the objective function of theoptimal transformation matrix is constructed according to the combination of the LPPalgorithm and Shannon Entropy. Finally, we minimize the constructed objectivefunction to find the optimal transformation matrix and the classification information.Compared with the local linear manifold feature extraction algorithm, the analysis andface recognition experiments show that the classification performance of featureextraction is improved by keeping the data of local features with the proposed algorithm.When the number of training samples is6, the average recognition rate of LPPalgorithm are respectively94.2%and78.5%for ORL face database and Yale face database while the average recognition rate of the proposed method are reached95.9%and87.2%respectively.On the other hand, the global and local information of data cannot be givenconsideration in2DLPP. In order to solve this problem, a bi-directional featureextraction algorithm based on locality and globality, which is abbreviated (2D)2PCALPP,was proposed in this paper. Based on2DLPP, the proposed method improves theperformance bi-directionally by fusing with two dimensional principal componentanalysis (2DPCA). Thus, not only the data after the dimension reduction can be overallreconstructed and represented but also the locality relationship can be maintained.Meanwhile, the classification information is also considered to improve theclassification performance. Relative to the2DLPP and2DPCA, the analytical andexperiment results show that the maximal recognition rate and average recognition ratecan all be increased. When the dimension of feature vector is80×2in FERET facedatabase, the maximal recognition rates are respectively80.73%and81.96%for2DLPPand2DPCA, while it reached84.21%in proposed method with the lower dimension offeature vector i.e.,9×9. As the number of training samples is5in the Yale face database,the average recognition rate of the2DLPP and2DPCA are respectively87.45%and89.36%, while it increases to91.25%with the proposed method.Against the inadequate of the typical locality feature extraction LPP and2DLPP,this thesis has been researched the feature extraction algorithms including the localitypreserving projections based on Shannon Entropy for feature extraction, and thebi-directional feature extraction based on locality and globality. Relative to the existingLPP and2DLPP methods that is the typical manifold learning algorithms, the researchesextend their contents and have profound theory significances and practical values. Theresearch results can be adopted by the existing methods of face recognition.