Study On Vehicle License Plate Recognition Technology

Automatic recognition technique of vehicle license plate lies core position in ITS(Intelligent Traffic System), which is one of deciding factors that contribute to automatic run of ITS. Among automatic vehicle license plate recognition techniques, the one based on image was a comparatively advanced technique, and was also the main direction of present research about automatic vehicle license plate recognition technique. However, technique based on image involves numerous subjects and lacks guidance of a uniform theory, there were too many disturbing factors in image, the contradiction between algorithm’s real-time performance and accuracy was difficult to solve. So, technique based on image was also the most difficult.In this thesis, in-depth, comprehensive, systemic analysis and study about the research status quo and achievement of vehicle license plate recognition technique based on image have been done. After that, this thesis pointed out the premise to enhance vehicle image’s grayscale contrast, picked out a binarizing algorithm(Otsu algorithm) not only has better effect, but also can elimilate background disturb to a certain extent, flexibly applied Sobel algorithm to edge detecting aiming at successive vehicle plate locating algorithm, improved the edge point density method and the vehicle license plate locating method based on grayscale break point, pointed out the limitation of applying Hough transform to rectifying sloping vehicle license plate and proposed a simple and practical rectifying method, proposed a character segmenting mehod based on the third character, used 13 feature as classifying feature of alphameric character and extended 13 feature to 22 feature, then used it for Chinese character classifying feature, done a comprehensively contrastive analysis about present representative neural networks, then used BP neural network to identify character and designed two BP neural networks to identify the Chinese character and the alphameric character respectively, done a comparatively in-depth study about an algorithm(i.e. Deterministic Perturbation Algorithm) used for BP neural network trainning and pointed out the shortage of this algorithm, finally, used BP algorithm with momentum for BP neural network training and the momentum was removed and introduced in terms of the network’s output error during training. In addition, some little improvement have been made on the other processes belong to recognition procedure.At the same time of theoretical study, all necessary processes of the whole recognition procedure have been implemented in C++ language on the platform of Microsoft Visual Studio IDE, the main body code and testing code totaled to 12621 lines, the data of every process’s technique guide line have also been presented by experiment. Experimental result shows that the algorithms improved in this thesis have comparatively satisfying effect, the rectifying method proposed in this thesis possess anti-jamming capability to a certain extent and can rectify vehicle license plate whether has edge frame or not, the character segmenting algorithm proposed in this thesis can meet requirement to a certain extent, the neural network designed by this thesis has reasonable structure and its system cost was not large, its training algorithm was effective, can meet the requirement of recognizing vehicle license plate characters.At present, there existed many pratically commercial license plate recognition products come into operation, but most of them aimed at simple scene and require many spot conditions to be met. Many algorithms proposed by this thesis possess practical value to a certain extent and can be used for actual license plate recognition system after appropriate modification.