Study On Image Classification, License Plate Recognition And Application On Embedded System Based On Support Vector Machines

Small sample learning problem can be solved by SVM very well. The aim of learning is not only to get optimal values when samples tend to be infinite, but also to the get the optimal solutions under current conditions of information. When samples are linear non-separable, the input vectors are mapped to a high dimensional feature space using the predefined nonlinear mapping, and then one optimal hyper-plane is constructed in the high dimensional feature space. It mainly depends on the kernel function that is used by SVM and some complicated computing can be avoided in the high dimensional feature space. In this paper, the geometrical model of SVM classification is shown and C-SVM, v-SVM classification methods are analyzed theoretically. Then several SVM classification algorithms are discussed. Specially, SMO algorithm is studied deeply.On the basis of analyzing Multi-classification SVM, aiming at the question of a small quantity of image classification, we present a new Multi-SVM classification method based on MLP and unilateral binary decision tree. Through learning of MLP network, Lagrange multiplier vector and threshold value b in decision function, constant in kernel function, restricted value in C-SVM and v-SVM classification can be gotten. In the end, images are classified step by step using unilateral binary decision tree. When classifying, one procedure of adjusting parameter is adopted to shorten the time of classification and advance the precision of classification.For actual application, especially for locating license plate in car plate recognition system, firstly, from viewpoint of texture image classification, angle second moment, contrast, relativity and entropy of image are gained to constitute one vector in spatial domain. Then the primary car image is transformed using DCT, and qualification value of DCT coefficients is gotten. From these values, the statistical and directional features of image are picked up to constitute another vector in frequency domain. At last, the car image is classified for locating license plate with two steps: coarse classification and fine classification. When classifying, thinking of speed, precision, area of interest, a flexible plan can be made. It is a good and operable method for classifying texture image such as car.For character recognition in car plate recognition system, the input vector parameters of SVM are mainly constructed from edge feature of image and projectivefeature of character. Thinking of the particularity and significance of edge, an edge detection method based on edge growing is presented for gray image and an edge detection method based on wavelet and reduced dimensionality model of RGB is presented for color images. These means assure we obtain the edge information as many as possible and avoid the bad influence of wrong input vector parameter. We present a method to construct image feature vector using edge and projection of character image. The composing way and normalizing of feature vector are shown. Finally, a good result is obtained when character is classified using unilateral binary decision tree.For application of license plate recognition, A real-time, partitionable task embedded system is studied and designed. The system includes image acquisition, image recognition, image transmission, device controlling and so on. With the strong parallel computing capability of C6000 series DSP of TI company, robust real-time controlling capability of StrongARM of Intel company, and flexible temporal logic controlling of FPGA, the system can achieve a high speed.The real-time operating system DSP/BIOS, which is used in image processing module, is analyzed and some key problems of DSP/BIOS are studied detailedly. Then Linux operating system, which is used in StrongARM real-time controlling and transmission module, is studied, mainly including real-time characteristic, memory management and device driver development. In the end, the main realizing means of developing car license plate recognition system are shown.