Research On Feature Fusion And Biomimetic Pattern Recognition For Biometric Recognition

As a typical application of pattern recognition, image recognition is to process, analysis,and understand image via computer, to recognize all kinds of patterns. As the basictechonology of digital image processing and artificial intelligence, image recognitiontechonology has been a hot research topic in pattern recognition and image understanding forpast two decades. It recognizes image pattern via image feature, and has been widely used inbiometric recognition (face recognition, finger print recognition, finger-knuckle-printrecognition, palmprint recognition and iris recognition), and handwriting recognition etc. Butthe robust image recognition in complex condition is still a difficult research problem. Thisthesis mainly studied the implementation and robust image recognition under complexconditions. The thesis focuses on the hot isssue of biometric recognition, and study thecommon and differeciential problem on biomimetic pattern based image recognition and theeffective feature fusion method, aims to propose the more suitable high-dimensional coveragemethod of the same class sample and the application of biomimetic pattern on featurerecognition.Specifically, the main work and results of this dissertation can be summarized as follows:1. A local feauture fusion approach towards disguised face recognition is presented.According to biomimetic pattern and information fusion theory, a disguised face recognitionalgorithm based on local feature fusion and geometry coverage via none disguised modelingstrategy is presented in this paper. Local binary pattern (LBP) and local phase quantization(LPQ) is firstly applied to extract the binary and phase statistics features which are robust tothe disguised mode, then hyper sausage neuron based on biomimetic pattern recognition (BPR)theory is adopted to construct high-dimensional geometry coverage of different classes, whichmakes full use of continuous characteristics of identical class face features while avoids theinterruption of the disguised mode. Experiments on AR face database and disguised facedatabase established by police face combination software show that, compared with thestate-of-the-art methods, the proposed recognition algorithm can achieve high recognitionresults under disguised conditions.2. A novel multi biometric fusion recognition modal based on iris and facial featurebased on biomimetic pattern recognition is constructed. Fusion biometric recognition modalcontributes in two aspects. The Contourlet transform (CT) and two directional twodimensional principal component analysis,(2D)2PCA, are used here to extract the iris feature and the facial feature respectively, and a new fusionfeature vector was formed on thecombination of the previous iris and facial features. Lastly, the fusion feature vector is used toconstruct the covering of high dimensional space using biomimetic pattern recognitionmethod, in which the hyper-sausage neuron is adopted. Furthermore, a principal componentanalysis based dimensional reduction method is adopted here to reduce the computationalcomplexity and improve the recognition efficiency. Experiments on the public union databaseshow that the proposed modal can achieve the state-of-the-art recognition accuracy whilekeeping the enrollment process safe. It can not only improve the biometric recognitionaccuracy, but also gives a comparatively safe strategy, since it is difficult for intruders toachieve multi-biometric information simultaneously, especially the iris information.3. After the analysis of multispectral effect for palmprint recognition, BP neural networkand convolutiaonl deep belief network (CDBN) multispectral palmprint image fusion andrecognition algorithm is presented. Firstly, a Harr wavelet based multispectral palmprintimage fusion method is presented. This method uses wavelet decomposition low frequency asfusion strategy, and then block singular value decomposition (B-SVD) and BP neural networkis adopted for feature extraction and recogniton seperately. Experiments show that theproposed fusion method can contribute to recognition results and achieve better performance.Futhermore, the deep CDBN network based multispectral palmprint recognition is alsopresented here. The unsupervised way is adopted in this method to extract the deep feature,which doesn’t rely on the common feature extraction method and can achieve betterrecognition results.4. A finger-knuckle-print (FKP) recognition algorithm based on image set and convexhull optimization model is presented. The presented algorithm first finds out the suitableconvex hull optimization model, and give detail analysis on the construction and optimizationof the model; then image set is used as the biometric input and local phase quantizaition isadopted for feature extraction in order to finish finger-knuckle-print recognition. Simulationexperiments show that, the proposed image set based recognition algorithm can acheive goodperformance on the public FKP database.5. Real-time face recognition system is designed. Face detection in the presented facerecognition system uses the YIQ color space to divide the skin color, find the eye centroid tolocate, adjust the misalignment or rotation, and finally output the detected face image;meanwhile, LBP method is adopted to extract feature, and the improved hyper neuron basedbiomimetic pattern is utilized for the recognition model. The system can achieve relative goodreal-time and good recognition results on cooperated conditions which approve the feasibility and effectiveness of biomimetic pattern reocognition algorithm.