The Research Of Similar Face Retrieval Based On SimHash Algorithm

In recent years, content-based image information retrieval technology is becoming ahot research topic. At the same time, as a contactlessly and friendly identity authenticationtechnology, there are more and more application requirements related to the facerecognition and retrieval technology. This thesis has made a research on the retrievalmethod under mass data based on the face object, and the main work includes thefollowing aspects:First, in the aspect of face detection and recognition, the thesis mainly discussed therapid face detection method based on AdaBoost, and the classical face recognitionalgorithm based on principal component analysis, and then applied both of them to thissystem. AdaBoost method has many merit such as high speed of calculation, high rate ofdetection, and strong robustness of attitude varying and blocking of faces, and isparticularly suitable for the real-time system application; The PCA method project thestandard faces onto the feature space formed by principal components, it emphaticallydistinguish the differences between different faces. PCA is realized to accomplish the firstdimension reduction in the system.Second, in the aspect of face retrieval, this thesis introduced SimHash method andtried using both PCA and SimHash to reduce the dimension of face image twice. In viewof the poor performance of traditional face retrieval technology when confrontedwith high-dimensional vectors, SimHash method can avoid the "curse of dimension" andsolve the problem that the speed of retrieval decreased a lot when the dimension of featurevectors become higher. First of all, extract the feature vector of face image, removeredundant information and get the principal component feature vectors using PCA, that thefirst dimension reduction. And then, SimHash is used to make a second dimensionreduction. The hash values resulted by SimHash are required spreading out to232quadrants as even as possible, And the final32-bit sequences are obtained as facesignature at last.This thesis finally make a simple implementation and validation of the method atWindows workbench. The system is set up with the use of the Virtual Studio2008, MFCframework and Opencv library. This thesis made a brief evaluation in the performance ofthe proposed algorithm based on ORL face database using Matlab. As you can see, due to the influence of the hash conflict, the accuracy of face retrieval algorithm is a slightdecrease in the situation than direct PCA retrieving, but improved on the average time ofretrieval system.