Statistically Effective Compact Local Binary Pattern For Face Recognition

With the rapid development of technology,how to protect the security of information is essential to each citizen and then the technology of biometric becomes the bright star of the times.As a member of the biological features,face feature attracted many researchers because of its unique advantages over the years.Today,face recognition has been widely used in access control systems,attendance,video surveillance,criminal investigation,etc.Many excellent algorithms for face recognition have been proposed,and the performance of face recognition systems has been improved gradually.However,under current circumstances,the recognition rate of most face recognition systems is still unsatisfied because of the existence of many uncertainties,such as light,face gestures,shelters,etc.Therefore,in order to improve the stability of the face recognition system,the research on the face recognition technology still has some theoretical significance and practical application prospects.Local feature descriptors,such as Gabor and LBP,are a kind of excellent facial feature extraction algorithms.LBP and its improved operator have been widely used in face recognition system because of its strong discriminative ability,insensitivity to illumination,and so on.However,LBP operator still has many defects,as follows.First,it needs to be treated according to predetermined rules which require strong prior knowledge;Second,it is extremely tactful to noise;Lastly,the information extracted is incomplete because that It can only acquire the sampling points in the fixed neighborhood.To cope with these problems,the main research works are described as follows:1.We propose a DLBP feature learning method for feature extraction.For each face image in the training set,the pixel difference matrix(PDM)of each position is extracted.Then,we learn a feature mapping to project each sub-region path vector into a gray value where 1)the variance of the same person is minimized,2)the variance between different people is maximized.Compared with the original MB-LBP operator,the algorithm is more discriminative.In order to reduce the complexity and improve the computational efficiency,the principal component analysis(PCA)is used to reduce the dimension of the extracted features.2.We propose a CLBP feature extraction method for face representation.The DLBP operator needs to extract the pixel difference vector of each sub window in the neighborhood window resulting in higher computational cost.For each face image in the training set,the pixel difference vector(PDV)of each position is extracted.we learn a feature mapping matrix to project each PDV into lower dimensional binary vector.The matrix should satisfy the following constraints: 1)the variance between each binary vector is maximized,2)the variance between binary vectors and original PDVs is minimized.3.We propose a SE-CLBP feature learning method for face representation.In this method,we use the histogram to count all the binary vectors which are converted by CLBP operator,and select N binary patterns as the main coding pattern according to the percentage.Finally,experiments on FERET face database show that our methods outperform state-of-art face descriptors.