Face Recognition Approaches Based On Circularly Symmetrical Gabor Tansform And Sparse Repressentation

As a research hotspot of artificial intelligence and image processing,face recognition technology has been applied widely in recent years.It is an interdisciplinary subject involving pattern recognition,image processing and artificial neural network,which has profound theoretical significance and extensive application.With the development of information technology,the requirement of the society to the technology is higher and higher.A good face recognition system,not only has to identify the subject accurately,but also has a fast recognition speed.The difficulties in face recognition are mainly from the change of facial expression and pose,and the impact of light during image acquisition.In order to reduce the influence of these two factors on the recognition effect,feature extraction in face recognition technology is particularly important.It will be of great help to classification if we can quickly and accurately extract good features.In this thesis,CSGT is used to extract the features of face images.Compared with the traditional GT,CSGT extracts the feature data with much lower redundancy,which is very favorable for the calculation of latter processing.Moreover,because of the correction of the direction information,CSGT has strict rotation invariance,which makes the extracted features more robust to illumination,expression and pose.The amplitude information of the circularly symmetrical Gabor feature is fused according to the three feature fusion schemes.The new feature images are obtained and used.Firstly,the CSGTp-SRC algorithm is proposed.The amplitude feature of the CSGT domain is merged by the average figure method.Then,in order to reduce the dimension of the feature image,the PCA method is used.Finally,the SRC algorithm is used to classify the feature.In order to further reduce the running time of the algorithm,the CSGTp-KSRC algorithm is proposed,and the KNN method is used to choose K nearest neighbors of a testing sample from all training samples to represent the testing sample linearly.Experiments on AR face database and Yale B face database demonstrate the feasibility of the two algorithms proposed in this paper.Moreover,although the CSGTp-KSRC algorithm is not as good as CSGTp-SRC.the run time of the algorithm is reduced by half.In order to reduce the running time of the algorithm,the paper also proposes a face recognition algorithm based on CSGT and CRC.The amplitude feature images of the CSGT domain are fused according to two fusion schemes.One is to connect each person's five amplitude images up and down to form a characteristic image,the other is to find five amplitude images of the maximum and minimum map,and then the two images are connected to form a feature image.The constructed feature images are dimensioned by PCA method,and then classified by CRC method.The algorithm is denoted as CSGTmagn-CRC and CSGTex-CRC respectively.The calculation speed of CRC is much faster than that of SRC,so the algorithm based on CSGT and CRC is faster.At the same time,KNN is used to further reduce the running time of the algorithm.The feature images of CSGT are constructed according to the two fusion rules,and the KNN method is used to choose K nearest neighbors from all training samples to compose a new set of training samples,and then the CRC algorithm is used.The corresponding algorithms are denoted as CSGTmagn-KCRC and CSGTex-KCRC.Four algorithms are analyzed by experiments in the AR human face database and the FERET face database.The four algorithms we proposed have higher recognition rate than the original algorithm,and CSGTmagn-KCRC and CSGTex-KCRC have a faster recognition speed.Through the experiments on the AR face database,the algorithms proposed and some basic algorithms are compared,and the feasibility of the algorithm we proposed is proved.The recognition rate and calculation speed of the algorithms based on CSGT and CRC are better than the algorithms based on CSGT and SRC.And KNN is used to further reduce the training sample dimension,although the recognition rate has declined,but the computing time of the algorithm lightly has been further reduced.