Research On Fast Algorithms For Face Recognition

The face recognition technology has been widely researched and used extensively.However,as the application in the field broadening,the target face library in face recognition task is more and more large,springing up vast amounts of face images,therefore,both quick and accurate method of face recognition retrieval is becoming more and more important.In the aspect of human face feature extraction,through the summary and analysis of existing methods,we choose the local feature extraction algorithm LBP having good performance for face recognition.the LBP based on block can more reflect the details of face features which is more robust to illumination changes,so we make sub-block processing for LBP images.Moreover,because feature dimension of each block is higher and contains redundant classification information,this article makes use of the combination of PCA and LDA to firstly achieve dimension reduction for it,then putting in series with the features of each block and getting the final face feature information.On the basis of above work,in view of the traditional facial recognition retrieval methods confronted with huge amounts of high-dimension data,usually facing "dimension disaster"and leading to querying slowly,the article focus on the rapid retrieval hash algorithm for high dimensional feature.Firstly introducing local sensitive hash algorithm and iterative quantitative sensitive hash algorithm(ITQSH),however of which the threshold of hash code selecting for zero is not an optimal value,the article proposes optimally looking for the optimal threshold method one by one for hash code,and designing the best classification indicators in class and between class in the process of the training samples hash indexing and binary coding.At the same time in the process of building the new hash table,by using boosting algorithm,increasing the weight of samples classifying falsely makes different hash tables possessing complementary classification effect.Finally we carry out experiments on extracted features of LBP based on block in the FRGC2.0 face database.It turns out that the algorithm putting forward is superior to the other on the indicators of face recognition preferred and querying time online.