Statistical Approach Of Fuzzy Vehicle License Recognition

License Plate Recognition is an important part in modern intelligent transportation systems,and car license plate recognition system contains three main components, namely the licenseplate location, character segmentation, character recognition. In this thesis, the two-dimensionallinear discriminant analysis (2DLDA), weighted modular2DLDA, incremental EMalgorithm, two-dimensional Fuzzy C-means clustering (2DFCM), increasing2DFCMalgorithm are used to study the license plate recognition, character recognition especially.Details are as follows:1. The thesis analyzes the advantages of the two-dimensional linear discriminant analysis(2DLDA) and the maximum scatter difference discriminant criteria, by compared with thetraditional LDA.2DLDA using a2D image matrix to calculate, let image separability in thefeature space that composed of less feature vector can obtain the best. The algorithm solvesthe small sample size problem which appears when using LDA in feature extraction.2. The primary task of establishing the Gaussian mixture model is to determine the numberof ingredients that comprise the hybrid model. To be able to automatically determine thenumber of ingredients according to the different input models, the thesis introduces the ideaof incremental application of the EM algorithm to establish the Gaussian mixture model. Theexperiments show that the proposed algorithm can adaptively determine the number ofingredients and achieve the coarse clustering.3. FCM algorithm has many shortcomings. An initial center will get different clusteringresults and affect the stability and accuracy of clustering, so it is the most important to select.The thesis introduces the incremental2DFCM algorithm. The experiments show that theproposed algorithm can adaptively determine the number of ingredients and the segmentationresult is satisfied.4. Modular2DLDA can not highlight areas that can distinguish the type of the characters.The thesis proposes a new method of weighted modular2DLDA combined with the idea of theRelief algorithm. In this algorithm, the image is segmented as needed into different sub-blocks.Being given the appropriate weights by the Relief algorithm in each sub-block, it has a differenteffect on the classification and recognition. The experiments show that the proposed algorithmenhances the anti-interference ability.5. In the thesis, three kinds of algorithms will be used to experiment on the same set ofcharacter images. The experiments show that the improved algorithm has many advantagescompared with the other algorithms. Last, the thesis analyze the relationship between thecharacter recognition rate and the character defect rate, as well as the character misrecognitionreasons.