Research Based On Feature Extraction For Face Recognition

With the development of computing and information technology, face recognitiontechnology become a mature technology, which has been used widely in various fields.Each of these recognition methods and algorithms has its strengths and weaknesses. Thethesis investigates some algorithms of feature extraction in face recognition, andproposes several optimized methods. Experimental results show that our methods aremore effective. This thesis proposes following three methods, as follows,①Feature extraction algorithm based on NMFNon-negative matrix factorization is a new sub-space analysis method. It appliednon-negative constraint to the pixel and reconstructed coefficient of basis Image, andmakes the reconstructed image is superimposed on a combination of non-negative inbasis image. It much more fulfills the concept of “constitute the whole of local” inhuman thinking. The basis image calculated by NMF is non-negative, which canovercome the local deformation of the face image. It is based on local characteristicsextraction, and the result has an actual physical interpretation. As with non-negativeconstraints, a face image is decomposed into non-negative linear combination ofnon-negative basis matrix, which can make the linear relationship more significant afterimage dimension reduction, then we can use linear regression classifier forclassification.②Feature extraction algorithm based on2DNMFThe original improved algorithms add some other restrictions based on two types ofloss function for non-negative matrix factorization(loss function on Euclidean distanceand K-L Dispersion), such as N2DPCA, which added restriction on principal componentanalysis based on the loss function of K-L. The traditional improved algorithms basedon two-dimensional non-negative matrix decomposition is not only to calculate the basematrix, but also required to calculate the coefficient matrix, and has a highcomputational complexity for the big dimension of the coefficient matrix. This thesisadds non-negative constraint to the two-dimensional principal component analysis,which makes the feature extraction process does not have to calculate the coefficientmatrix, just extract the base matrix. For example, there are m training face images sizedp*q,2DPNMF simply extract base matrix sized p*r. In conclusion, it is very simple in2DPNMF iterative process, it greatly reduce the computational complexity and featureextraction time. We did experiments on Yale, FERET and AR face library, and the result shows that our algorithm combines the advantages of non-negative matrix factorizationand two-dimensional principal component analysis, not only the recognition rate hasimproved greatly, but also the speed is significantly better than the non-negative matrixfactorization, or even slightly faster than the two-dimensional principal componentanalysis.③Feature extraction algorithm based on B-spline and Image gradientIllumination change is an important factor to affect the accuracy of the algorithm inface recognition technology. In recent years, this problem has been extensive researched,and can be divided into the following categories. The first one is image normalizationby image processing technology under different light conditions, such as histogramEqualization and so on. It is very difficult to calculate the different light conditions Inthe actual scene. The second method is to build a3D face model to solve the lightchanges; however, to build a3D face model requires a lot of face training picture. Thethird method is extracted the illumination invariant feature from the face image, but thismethod is very unstable for the recognition rate of the noise image. For these problems,this thesis proposed a single sample identification technology based on B-splinefiltering and insensitive gradient face lighting. As extraction feature in gradient space,this method is very sensitive for the effects of noise. Traditional denoising methods,such as Gaussian filtering, the filtered image will be too smooth, it is not impact greatlyfor recognition algorithm based on the pixel space, but very significant for therecognition algorithm based on gradient space, which will lose some of the edgeinformation, and the approximation of original image became deteriorates. For theseproblems, this thesis proposes a single sample identification method based on B-splineand gradient face, the algorithm choose B-spline function with a low-passcharacteristics as a smooth function. B-spline function is order tunability, when theorder value is large; it performs better in smooth, which can make good noisesuppression. When the order value is small, it performs better in approximation, whichhas a better approximation of the original sample.