Research On Methods For Image Feature Description Based On Local Binary Pattern

The extraction of effective features is a key problem in digital image processing and computer vision.In recent years,feature extraction methods based on local binary pattern(LBP)have been widely applied to the fields of texture classification,scene categorization,face recognition,image retrieval and pedestrian detection,and so on.However,existing LBP-based methods have problems in extracting image features,such as sensitivity to inverse gray-scale changes and insufficient descriptive ability for color image description.To solve these problems,the following research work is carried out in this thesis:1.The existing LBP-based methods are sensitive to inverse gray-scale changes.To mitigate this problem,a feature extraction method based on sorted local gradient pattern(SLGP)is proposed.Firstly,two complementary local gradient patterns(LGP)are proposed to encode rich gradient information present in a local neighborhood.Then,the dominant intensity order measure(DIOM)is proposed to sort image pixels into two categories,followed by extracting LGP features over the categorized pixels.In this way,SLGP encodes both local gradient information and global gray order information and improves the classification accuracy of texture images under inverse gray-scale changes.2.As existing LBP-based methods are sensitive to inverse gray-scale changes,two feature extraction methods based on complement coding are proposed,i.e.,completed local complement and gradient pattern(CLCGP)and local complement and derivative pattern(LCDP).In CLCGP,the signs of neighbor differences neglected by SLGP are encode by a rotation invariant local complement pattern(LCP).The final CLCGP descriptor is built based on a 3-D joint histogram of the signs of neighbor differences,the magnitudes of neighbor differences and the intensity of center pixels.In LCDP,a local derivative pattern(LDP)is introduced to encode the magnitudes of neighbor differences in(the first and the second order)Gaussian derivative space at different scales.Furthermore,a joint coding scheme based on mean sampling is proposed to consider both difference signs and difference magnitudes information.Finally,the LCDP descriptor is obtained by constructing multi-scale histograms representation of jointly encoded features.Experiments show that the CLCGP and LCDP methods can effectively improve the classification accuracy of LBP-based methods under both linear and nonlinear inverse gray-scale conditions.3.The existing LBP-based methods have weak description ability for color images.To improve these LBP-based methods,a feature extraction method based on quaternionic extended local binary pattern(QxLBP)with adaptive structural pyramid pooling(ASPP)is proposed.Firstly,an extended quaternionic representation(EQR)is proposed by introducing an information term as the real part of a quaternion.The resulting EQR can flexibly encode some discriminative features and handle multichannel images.Then,a QxLBP descriptor based on LBP is proposed to encode local neighboring information and complementary modulus and phase information in the quaternionic domain of color images.Finally,a multi-resolution ASPP method is proposed.Unlike the traditional spatial pyramid pooling(SPP)which is sensitive to image rotation and spatial changes,ASPP is a structure-oriented feature pooling method which can adaptively aggregate the encoded features according to the multi-resolution characteristics of image.Experiments for color texture classification and scene categorization demonstrate that the proposed method improves the description ability of existing LBP-based methods for color images.