Research On Face Feature Extraction Technology Under Complex Illumination

Face recognition technology has become one of hot topics in the field of computer vision,and it is widely used in the fields of Face ID,security-monitoring,criminal detection and more.In real application,although conventional face recognition methods have achieved good results in specific scenes,it is affected by illumination,expression,posture and occlusion,which lead to sharp decrease in recognition rate.Among the factors,complex illumination is one of the most significant impacts on face recognition.In order to obtain insensitive-illumination image feature,this paper improves several local feature extraction algorithms.The main work and innovations in this paper are shown as follows:(1)Focusing on the drawback that Weber Local Descriptor(WLD)operator can't fully describe the local texture details,a novel method named as Synergistic Weber Excitation Patterns(SWEP)based on optimal selection is proposed.Firstly,an optimal neighborhood structure based on synergistic center-surround receptive field model is established to provide richer candidates structure.Then,the feature optimization selection method based on improved mutual information is used to learn SWEP according to criterion of minimizing redundancy and maximizing relevance,next,DSWEP codebook is generated.DSWEP can make full use of class information and adopt the inner and outer two-layer neighborhood structure,which not only represent illumination-insensitive feature effectively,but also has strong robustness.(2)Focusing on the difference of SWEP neighborhood structure candidate model for classification,a feature optimization learning method based on Boosting for DSWEP is proposed.This method utilizes different weight coefficients base on the descriptive ability of features to generates C-DSWEP features,this method improves the discriminative ability of DSWEP features.(3)Focusing on the shortage that Local Binary Pattern(LBP)operator is sensitive to noise and can't adequately reflect the gray level change of neighborhood,a Centrosymmetric Neighborhood Weighted Average Local Binary Model(CNALBP)is proposed.Firstly,the traditional LBP single-layer neighborhood model is extended to the two-layer neighborhood model,and then the two-layer neighborhood pixels in each direction are weighted averaged.Finally,the weighted average of the neighborhood about the symmetry of the central pixels are compares and codes them according to certain rules.The algorithm can effectively reduce the computational complexity and feature dimension.Base on this,combining the advantages of CNALBP and HOG operators,we propose an image recognition algorithm based on fusion of CNALBP features and HOG features,which has great robustness to illumination.In this paper,extensive experiments are conducted on face databases of CUMPIE,Yale B and FERET.The proposed improved algorithms are compared with the state-of-art feature description algorithms.The experimental results show that the proposed operators are robust to illumination changes.It has a certain academic value and broad application prospect.