Research On Content Based Handmetric Identification

The identity recognition is a long-lasting topic that relates to national security and society stabilization.Under the traditionally used identifier and password based identity recognition technology,the requirement of society security will not be perfectly satisfied.Therefore,the biometrics recognition technology is developed as being demanded.Along with the mature development of the technology,working environment of biometrics identification technology is gradually transformed from closed laboratory to complex and open.Recognition technology based on fingerprint and lace has already been widely adopted in many real-life applications.Compared with fingerprint and face recognition,palmprint based biometrics identification is a newly emerging technique and needs further development.Single palmprint features mainly include structural features,frequency domain features,subspace features,and statistical features.To utilize more available information,multi-module features fusion are gradually adopted.In all the above-mentioned feature forms,images are processed as a matrix more than an image,only with the structural feature extraction as an exception.Less attention has been paid on color,shape,texture,spatiality,regional visual information and image semantic information which are contained in images.Those information are called content features,they can represent the image content in different levels,which make the Content Based Image Retrieval(CBIR)becomes the advanced image retrieval technology comparing with the Text Based Image Retrieval(TBIR).This dissertation discusses the image content based handmetric identification,to solve the problems that emerge during the industrialization of the handmetric identification,like the identification of low quality hand images captured under open environment,or how to keep the system performance when very few images are available for training.The main research work of this dissertation includes the following aspects:1.Analyze the current handmetric identification algorithms,and find out the existing problems during the technology development and industrialization.Propose the generalized and narrow sense definition of handmetric,and first use image content in representing handmetric images to solve problems and accomplish the identification under different handmetric definitions. 2.For the translation and rotation that usually happen in image capturing,wavelet Multi-Resolution Analysis(MRA)is combined with the moment invariant to analyze the units with different scales that compose the texture contained in palmprint ROI images.The moment invariant of each decomposed sub-image is calculated as the description of texture in low-level content features because moment invariant is robust to affine transformations.3.In low resolution hand images with narrow sense definition,the traditional feature forms like hand lines,wrinkles,minutiae,and textures are not that easily and clearly extracted as the images under regular resolution.Considering the characteristics of the low resolution images,a hierarchical identification method based on the hand geometry and gray-level distribution is proposed:at the coarse-level,angle information of the triangles constructed by the lines segments connecting finger valleys is added as a complement to line-segment-only hand geometry to specify the relative position of finger valleys;at the fine-level,with the assumption that the hand lines,wrinkles, texture and skin color of different position within a hand image will lead to differentiable gray-levels,the gray-level and spatial information are combined together in the form of local gray-level entropy to describe the local gray-level distribution of handmetric images.4.Region is adopted to be the representation of low resolution handmetric images as the structural unit instead of the traditional unit like point and line,to alleviate the instability in point and line extraction.Develop the Simple Sequence Labeling segmentation method of the low resolution handmetric images under narrow sense definition,with the assumption that the gradient value of each pixel represents its gray-level changing rate.Then choose conditional regions which are steady in segmentation through region area constraint.Because distinctive lines and dense textures always have lower gray-level values than their surrounding areas,regions with lower average gray-levels are selected from conditional regions.The centroid coordinates of these regions are calculated as feature vectors.Finally,a Regional Spatiality Relationship Matrix is defined to measure the distances of feature vectors with different dimensions.Experiments show that the usage of region as the structural unit is robust to rotation and translation.5.Obviously,it is sophisticated to directly extract text based distinctive semantic features from a handmetric image.We introduce the latent semantic analysis(LSA) which is used in text analysis,into the handmetric identification.LSA can avoid the comprehension of natural language,and extract the relation hidden in data as the semantic concept by statistically analyze a large scale samples.In guidance of the LSA, handmetric images' local information is extracted as the local semantic.The Non-Negative Matrix Factorization(NMF)based LSA of palmprint ROI images is proposed,to describe the handmetric image in the semantic level.We have modified the traditional NMF to solve the performance instability brought by random initialization, and to reduce the computational complexity caused by factorization.The improved NMF is used to substitute the SVD matrix decomposition method in LSA to achieve the real-time LSA of the palmprint ROI images.6.Making the subspace methods used in palmprint identification as an example,we discuss the principal component selection in subspace construction and its influence on the system performance.Through respectively removing continuous and single principal components,we try to find out the principal component selection scheme when the system achieves the optimal performance.Experiments made on different component selection prove that selectively throwing off some principal components will further eliminate the basis vectors that contain the redundant and disturbing information, therefore effectively enhance the system performance.