Font Size: a A A

The Application Of The Embedded HMM Based On Distance Transformation In Personal Identification With Finger-vein Pattern

Posted on:2007-11-02Degree:MasterType:Thesis
Country:ChinaCandidate:D Y MaFull Text:PDF
GTID:2178360182996627Subject:Computational Mathematics
Abstract/Summary:PDF Full Text Request
This dissertation concerned with the application of theembedded HMM based on distance transformation in personalidentification with finger-vein pattern contains six partsincluding exordium.In the part of exordium, we have briefly introduced therecent development of finger-vein recognition, the applicationof HMM in pattern recognition and the reason that we usedembedded HMM (EHMM) in our proposed finger-veinrecognition method.In chapter 1, we first introduced the method named"repeated line tracking" proposed by Naoto Miura et al. whichis used in feature extraction of finger-vein pattern;then weproposed a method for further noise elimination based on thearea and the maximum span.In chapter 2, after introduced Markov chain and HiddenMarkov Module (HMM) that based on it we give the basicarithmetic of HMM.In chapter 3, we detailedly introduced the frame of EHMMthat used in our finger-vein recognition arithmetic, and thendiscussed its state space and observation space in our algorithm.In chapter 4, we discussed two problems that we met whenwe used EHMM in finger-vein recognition, that is the overflowof arithmetic and the initialization of our proposed EHMM, andthen we gave the solutions of these two problems.In chapter 5, we described the experiments we ran toevaluate the performance of our proposed algorithm, and finallygave a conclusion of this dissertation.The main results from the dissertation can be listed asfollow.1. The feature extraction of the finger-vein patterns from theoriginal images.First, we applied the "repeated line tracking method" toextract the finger-vein patterns from the nonuniform images.Then we proposed a method for further noise eliminationbased on area and maximum span. The details of this methodare described as below.Step 1: Mark the different white connected regions usinglabeling algorithm.Step 2: Calculate the areas and the maximum transversal spansof the white connected regions respectively.Step 3: Select the befitting thresholds for the area and themaximum transversal span to eliminate the whiteconnected regions whose area or maximum transversalspan has a smaller value.2. The frame of EHMMWe adopted EHMM in our finger-vein recognitionalgorithm, its frame is briefly described as below.The parameters of the main HMM:● N :The numbers of the states in the horizontal main HMM,called the numbers of the super states.● π :The initial super state distribution, π = {π i ,1≤i≤N},Where π i is the probability of the i th super statebeing the initial super state, that is π i = P (Q 1 =Λi|λ)and 0 ≤ π i≤1, 11∑==Niπ i.●A:The super state transition probability matrix,A = {a ij ,1≤i,j≤N}where, a ij = P(Q x=Λj|Qx?1 =Λi)and 0 ≤ a ij≤1, 11∑==Nja ij.● Λ:The set of the sub vertical HMM embedded in superstates, Λ ={Λ i ,1≤i≤N}.The parameters of the sub HMM:● N i:The number of the states in the sub HMM thatembedded in the i th super states. We defineS i = {s ki,1≤k≤Ni} as the state space of the sub HMMembedded in the i th super state.● π i:The initial state distribution of the sub HMM embeddedin the i th super state, π i = {π ki,1≤k ≤Ni}and 0 ≤ π ki≤1, 11∑==Nikiπ k.● A i:The state transition probability matrix of of the subHMM embedded in the i th super stateA i = {a kil,1≤k,l≤Ni}where a kil = P( qxy=sli|qxy?1 =ski),and 0 ≤ a kil≤1, 11∑==Nilia kl。● B i:The output probability function, B i = {b ki(Oxy)}, wherebk i (O xy)= P(Oxy|ski). In a continuous density HMM, thestates are characterized by continuous observationdensity function that is typically represented in termsof a mixture of Gaussian functions.3. The observation vectors based on distance transformationIn this dissertation, to overcome the infection caused bytoo much useless observation vectors produced by background,we adopt the observation vectors that based on distancetransformation. The details of the distance transformation aredescribed as below.Step 1: We define the original image as f ,scan the image fromleft to right, and then from top to bottom, obtain theimage g using the function as below.g ( m,n)= ??? min{ g (m?1,n)+01,g(m,n?1)+1}iiff ff(m( m,n,n))==2055where, f ( m,n) is the intensity of the pixel ( m, n) inimage f ,image g is initialized as 0.Step 2: Scan image g in an inverse order, and obtain imageh as followh ( m,n)= min{g (m,n),h(m+1,n)+1,h(m,n+1)+1}where, image h is initialized as 0.h is the image wewanted after distance transformation.
Keywords/Search Tags:Transformation
PDF Full Text Request
Related items