Reserch On Reconstruction And Recognition Of Remote Side Face

Most of current face recognition is based on close and front face, the facerecognition based on long-distance and side face is of small part, Long-distanceidentification is not easy to be found by people. However, target images of side face willhave low resolution when under long-distance conditions, and the details are not easilydistinguished, super-resolution reconstruction is required for side face image of a periodbefore recognition to improve the recognition rate.A new adaptive regularized super-resolution reconstruction algorithm is proposedbased on the research of super-resolution reconstruction and recognition, and side facerecognition of long-distance and recogniton prototype system are implemented in thispaper. The main research contents are as follows:(1) In terms of image registration, a non-rigid registration method based on thecombination of brightness consistancy and piecewise smooth model is proposed，andthe Gaussian pyramid is involved to implement sub-pixel motion estimation.Both thelinear transformation of the whole image and the nonlinear transformation of localimage are taken into consideration to make the registration results more accurate.(2) In terms of reconstruction, fuzzy entropy is introduced as the judgment for thechange Intensity of grayscale in image space within the framework of regularizationbased on normal form. Neighborhood homogeneous measurement of fuzzy entropy isused as the weighted coefficient of regularization item in this paper to better recognizethe edge and smooth area of image, and different regularization weights are used indifferent image regions, which are significant for maintaining sharpening in edge regionand smoothing in flat area.(3) In terms of reconition, the combination method of Two-Direction Image MatrixBased Principal Component Analysis (2DIMPCA) and Multiple Discriminant Analysis(MDA) is used for the extraction of facial feature. Firstly,2DIMPCA is used for featureextraction and dimension reducing of side face image, and then original feature vector isobtained and vectored into feature vector; Secondly, MDA is employed on the featurevector of side face to obtain fusion feature vector; Finally, the recognition process isimplemented on the fusion feature vectors by Nearest Neighbor (NN) algorithm.Theresults of experiment show that the reconition of reconstruction face is of better effect. (4) A reconstruction and recognition system of remote side face is researched anddeveloped. Experiments are implemented on CASIA Dataset B database to prove theeffectiveness of reconstruction and recognition.