Research On Face Feature Extraction And Face Recognition Method Under Single Sample Condition

In practical applications,the collection of face images is limited by many factors,each person often only has one face image as a training sample,which results in less than ideal face recognition performance under single sample condition.Among them,the feature extraction method used in single sample face recognition has the problems of incomplete feature expression and high computational complexity.In addition,the virtual sample generated by the sample expansion method for solving the single sample problem is single and the correlation with the original image is too high,so that the recognition rate is low.For solving these problems,This paper studies face recognition under single sample condition from two aspects of feature extraction method and sample expansion method.Based on the research of Histograms of Oriented Gradients(HOG)feature extraction method,this paper proposes a face feature extraction method based on WPD-HOG pyramid;and proposes a face recognition method based on NMF to reconstruct virtual samples for the sample expansion problem.The main contributions include the following aspects:1.Firstly,a feature extraction method of WPD-HOG pyramid was given.The best HOG parameters are obtained through experiments by analyzing The influence of various parameter choices of HOG method on the face recognition rate in detail.The method combines the HOG,Image Pyramid and Wavelet Packet Decomposition(WPD)together to characterize the face image feature.The WPD-HOG pyramid features are identified by the SVM classifier for face recognition.This improved method was applied to single sample face recognition.Experiments are conducted over ORL data set to demonstrate the proposed approach.Compared with the benchmark methods,the experimental results show that the method proposed in this paper is suitable for face recognition in general situations and is suitable for single sample face recognition.In addition,the recognition performance,computation complexity and noise robustness of the proposed method are the best.2.Secondly,the paper studied the validity of the mirror expansion,sliding window method and bit plane method of the sample expansion method.To directly solve the problem of lack of training samples for a single sample,use the sample expansion method to generate virtual samples to increase the number of samples in the training set.Among them,the mirror transformation can enhance the robustness of the partial rotation of the image.The sliding window method simulates the different positions and distances of the lens and the human face when the image is sampled.The bit plane method gives the possible change of the pixel to a certain extent.Experiments show that the three methods all contribute to the improvement of the face image recognition rate,in which the mirror transformation has the best effect.3.At last,improved sample expansion method.Considering that non-negative matrix factorization(NMF)reconstructed face image can better represent the internal relationship between the local parts of the face,this paper combines NMF with mirror transformation,sliding window method,and bit plane method.Aiming at a single sample face,this paper proposed a sample expansion method for reconstructing virtual samples based on NMF.The effects of the number of base image r and number of iterations on the recognition rate of NMF are studied,and the best parameters are used to perform NMF reconstruction on the image.Single sample face recognition experiments on the ORL and FERET face database verifies the effectiveness of the proposed method.It can retain the information that contributes more to the recognition rate,removes information that is useless and interferes with the recognition rate,and improves the face image recognition effect.