Sparse Representation Of Face Recognition Method Based On Gabor And HOG Features

Face recognition as an effective biometric identification technology has become a hot issue in face recognition fields,and simultaneously obtained wide application in people's lives.In recent years,various kinds of algorithms appeared,such as local feature analysis based method,Eigen-face algorithm,the elastic model based method,neural network approach.However,in practice,these methods are easily influenced by the surrounding light;the result is also sensitive to the face posture,which is a huge challenge to deal with these conditions.Compared with methods mentioned above,sparse representation based face recognition method show high recognition rate and robust,it has become a hot topic in face recognition field.But there still exist some shortcomings such as large computation consume and lacking robustness for postures change.Gabor filter response mechanism can well reflect the human visual system,also can effectively extract local spatial and frequency domain features of face images;HOG features as a feature descriptor have obtained excellent performance in pedestrian detection fields,which is robust to light and posture changes.Combining these features,this paper presents a sparse representation based face recognition method fusing with Gabor and HOG features.The main work is as follows:1)First,Haar feature based face detector is introduced.After detecting face regions,we firstly get the position by detecting the eyes,then determine the inclination angle binocular connection,executing affine transformation algorithm to rectify the face region.2)By two-dimensional Gabor wavelet transform,we can effectively extract the local frequency and spatial features of the human face images,and simultaneously can effectively extract HOG(Histogram of Oriented Gradient)feature from the key places in face images such as nose region and mouth region etc.PCA(Principal Component Analysis)is employed to reduce the dimensions of the two features introduced above.A new feature is obtained by cascading the two low dimension features,which is employed for later face recognition algorithms.3)Over-complete dictionary is constructed by extracting the new feature descriptor introduced above.The paper also discusses the question that how to get the sparse solution based on defined norm.According to the reconfiguration-error,the paper proposes a self-updating dictionary method,which can effectively improve the face recognition rate.