Oriented To Confirm The Task Of Face Recognition Technology

Automatic Face Recognition (AFR) includes two tasks: identification and verification.It has received large progress in recent years. However, most research focuses on faceidentification. Only a few papers concern the verification of faces, yet have little deepresearch on its special characteristics. This thesis deals with the face verification. The mainwork of this thesis is as follows:(1) The specialty of face verification techniques is investigated in this thesis. Wecompare face verification with face identification from several aspects, and deduce severalmathematical relations among the Rate of Top One Match, the Equal Error Rate, and theMinimum Total Error Rate. This thesis presents a mathematical model to compute theidentification and verification performance, and uses this model to theoretically prove thatthe performances of identification and verification follow different rules of change when thedistribution of intra and extra class similarity scores change. It suggests that an algorithmwith good identification performance does not imply good verification performance, andvice versa.(2) The thresholding approach in verification systems is studied. To solve the "scorevariation" problem caused by appearance variation of images, this thesis investigates twomethods, Subject Specific Threshold (SST) and Score Normalization (SN). Throughtheoretical analysis and experiments, we prove that SST and Z-Norm methods can getapproximately equivalent effects in case that every subject in the system has only onetraining image. We perform experiments with multiple methods in both the FERET andCAS-PEAL face databases. The results reveal that SN can improve the performance of faceverification system in most cases compared with the unified threshold method.(3) The SVM based face verification algorithm is investigated in the thesis. We firstlycompute the difference vectors in PCA and LDA subspace, then use SVM method toclassify intra and inter class difference vectors. The SVM method is compared with severalbaseline methods using the FERET database. The results show that SVM method canimprove the verification performance in PCA subspace. However, it can not achieve betterverification performance in LDA subspace.(4) This thesis proposes a new face verification method based on Local Binary Pattern(LBP) and SVM. This method first extracts the LBP features from face images, thencomputes the local region similarities and concatenates all the local similarities into asimilarity vector. Finally, an SVM classifier is used to classify the intra and inter classsimilarity vectors. In all the probe sets of the FERET face database, LBP-SVM method canachieve better performance than non-weighted LBP algorithms for face verification.The research results in this thesis reveal that, although verification and identificationare similar in core techniques, verification has its unique characteristics. In addition, the"score variation" problem requires more attention.