Font Size: a A A

Study On Multidimensional And Multiresolution Biomimetic Recognition Method

Posted on:2012-10-03Degree:DoctorType:Dissertation
Country:ChinaCandidate:D WangFull Text:PDF
GTID:1118330335951992Subject:Agricultural mechanization project
Abstract/Summary:PDF Full Text Request
Biometric identification technology employs the inherent physiologic or behavioral characteristics of human being to recognize individuals. It uses many high-tech methods, such as computer algorithm, kinds of sensors, statistical algorithm and classification algorithm, to realize the identification. At present, biometric identification technology mainly deals with computer vision, image processing and pattern recognition, computer auditory, speech processing, multi-sensor technology, virtual reality, computer graphics, visualization technology, computer aided design, intelligent robot perceptual system, etc. The biological characteristics used in biometric identification include hand shape, fingerprint, face, iris, retina, pulse, auricle, etc., and the behavioral characteristics include signature, voice, keystroke, etc. Based on these characteristics, biometric identification technology has made considerable progress over the past few years.Since the beginning of the 1990s, professor Shoujue Wang has made many researches on traditional pattern recognition. He considered that the traditional pattern recognition uses "optimal separating" of different kinds of samples in the feature space as main principle, compares a new sample with the trained samples one by one, then judges the new sample whether belongs to some kind of trained samples. The traditional pattern recognition has two disadvantages:①It may lead the recognition error for the samples untrained in learning process;②When learning a new sample, the already trained knowledge base may be upset, which will destroy the recognition of the trained samples. To solve the above problems, professor Shoujue Wang proposed a new recognition method—Biomimetic Pattern Recognition (BPR), according to the objective law of recognition of human being. The theory of BPR provides a new idea for solving imaginal thinking problems. With the understanding of traditional pattern recognition, and in view of the shortcomings of traditional pattern recognition and the difficulty in constructing multidimensional space covering in BPR, a new biomimetic recognition method is proposed based on multidimensional space biomimetic informatics and multiresolution analysis in this paper. To verify the effectiveness, the method is applied to iris recognition. Experimental results show the proposed method is effective for pattern recognition application.(1) A new biomimetic recognition method is proposed based on multidimensional space biomimetic informatics and multiresolution analysis, which applies to many kinds of biometric identification. First, the multiscale features of samples are extracted by multiresolution analysis method and multiresolution robust representation at different scales are obtained by local histogram of oriented gradient descriptor. Second, the optimal multidimensional space covering of feature space is constructed by self-organizing map (SOM) clustering and distance projecting distribution in the feature space of same kind of samples. Finally, the sample to be identified is judged whether belongs to different scale covering sets to decide whether the sample comes from the same kind.(2) During the process of feature extraction and expression of biometric images, multiresolution analysis method is innovatively used to extract the features. In addition, to reduce the influences of collection angle, lighting condition, etc., Histogram of Oriented Gradient (HOG) descriptor is introduced to obtain the multiresolution robust representation of biometric images.(3) When constructing the optimal point-set covering, the center of same kind of samples is obtained by Self-Organizing Mapping (SOM) clustering algorithm in the high dimensional space, and it is taken as the centroid of hypersphere covering. The radius of hypersphere can be calculated by distance projecting distribution and "3σ" principle of probability statistics. Therefore, for each subband, the hypersphere covering can be constructed according to the centroid of same kind of samples and the radius of distance projecting distribution.(4) At recognizing stage, according to the theory of BPR, a sample is recognized by judging the sample vector whether belongs to the range of the connected point-set covering. If the sample is in the point-set covering, they will be the same kind. The same kind of samples are divided into many subbands in different feature spaces by multiresolution analysis. Hence, there are many point-set covering in the different scale spaces with different recognition performances. A kernel function is used to acquire the uniform recognition framework. By choosing a proper kernel function, the judge results of recognition in different scale spaces can achieve unification.(5) To validate the effectiveness, the method is applied to iris recognition. Iris images are easy to be disturbed by lighting condition, rotation, scale, translation. To obtain the stable representation of iris feature, the iris preprocessing including iris location, iris normalization, iris enhancement is performed and a 64×512 iris image was obtained. The multiscale features were got by GHM multiwavelets and HOG descriptor. The signal is decomposed by 1 level, which means 16 subband images will be obtained. The robust representation at each subband is obtained by local histogram of oriented gradient descriptor. Each block is 4×4, including 2×2 cells, and each cell includes 9 orientation binning in [0,π].120 iris images from the same person were used to construct the point-set covering by SOM and distance projecting distribution.16 covering were obtained in different scale spaces. Finally, the proposed method was validated based on JLUIRIS iris lib. The widely used ROC curve was adopted to measure the recognition performance. Results show the multidimensional and multiresolution biomimetic recognition method has better recognition performance.In summary, the multidimensional and multiresolution biomimetic recognition method applies to many kinds of biometric identification. It may provide the theoretical reference for other pattern recognition methods and researches on Multi-Dimensional Space Biomimetic Informatics.
Keywords/Search Tags:Multidimensional Space Biomimetic Informatics, Biomimetic Pattern Recognition, multiresolution analysis, multiscale, Histogram of Oriented Gradient, GHM multiwavelets, iris recognition
PDF Full Text Request
Related items