Font Size: a A A

Study On Face Image Feature Extraction Based On Local Binary Pattern

Posted on:2019-07-20Degree:DoctorType:Dissertation
Country:ChinaCandidate:J LiFull Text:PDF
GTID:1368330563992210Subject:Control Science and Engineering
Abstract/Summary:PDF Full Text Request
As an efficient image feature extraction method,LBP descriptor has been widely used for pattern recognition in recent years.Although a great deal of research has been done on LBP based face recognition,these works still leave much to be desired.This dissertation studies grayscale and color face images LBP descriptor,and further proposes five improved image feature extraction methods.The major research works of this dissertation are the following.As the existing Kirsch edge direction based LBP descriptors do not make full use of the direction information,a grayscale face image descriptor,local edge direction-texture descriptor(LEDTD),is proposed.In LEDTD,compute the Kirsch edge response of each direction and extract the minimum and maximum response edge directions firstly;then apply local binary coding strategy to extract edge direction based texture feature under the multi-direction framework;LEDTD is formed by fusing the edge direction and the texture information lastly.LEDTD differs from the existing edge direction based LBP methods in a manner that it not only considers image edge direction information but also extracts image edge direction based texture information.The existing first-order Riesz transform based LBP feature extraction methods only extract intrinsic one-dimension structure information.What's more,they do not consider the complement information among transform resultant.A multi-order Riesz transform based grayscale image descriptor,Riesz binary pattern(RBP),is proposed.In RBP,multiscale image analysis and multi-order Riesz transform are ultilized to extract image intrinsic structure information;RLBC is the local binary coding of local image structure feature extraction on each Riesz transform response;Making full use of the complement between different Riesz transform,RGBC is the binary encoding of the image Riesz transform joint information.RBP is the feature level fusion of RLBC and RGBC.LBP and its variant are all the encoding of integer order(e.g.first-order,second-order)derivative information between the center and its neighborhood pixels.As the generalization of the integer order derivative,fractional order derivative can extract more discriminant information for image representation.Therefore,proposes the completed local fractional order derivative feature vector(CLFDV),from which two image representation methods are proposed.One is clustering based image CLFDV vector quantization descriptor f VQP.Vector quantization based encoding method has better noise robustness than scalar quantization applied by LBP,and its space complexity is settable.The other is multi-level Fisher Vector aggregated multi-structure model based CLFDV higher order statistic image representation method m FVFD.m FVFD improves f VQP method from two aspects: it extract local multistructure feature;it is the higher order statistic of CLFDV.LBP is originally design for one channel image texture feature extraction.And an improvement of LBP,completed color local binary pattern(CCLBP),is proposed for color image representation.CCLBP is composed of two complementary components: soft color label and color local similarity code.Expressed by soft color label,clustering based feature vector quantization method and the soft-assignment coding method are used to summarize and encode the image pixel color information globally.While based on the central pixel and its neighborhood pixels color information similarity binary coding method,color local similarity code is used to encode the local spatial color textural feature.
Keywords/Search Tags:Face Recognition, Local Descriptor, Local Binary Pattern, Kirsch Mask, Riesz Transform, Fractional Order Derivative, Color Image Feature Extraction
PDF Full Text Request
Related items