License Plate Recognition Technology Based On Neural Network Research And Applications

License Plate Recognition(LPR) is a very important component of intelligent transportation system, which is the important application of digital image processing, artificial intelligence and pattern recognition in the field of intelligent transportation system. Therefore the study and the development on car license recognition system have the important value and the significance regarding our country traffic control domain's development.There are many ways to recognize license plate. BP neural network is used very much currently, but BP neural network has two shortcomings which are slow convergence and easy to fall into the local extreme point of the error function. To solve these problems, we have a study on the license plate character recognition with neural network architecture and genetic algorithm that has the characteristics of global optimal search capabilities. The main contents include the following three parts:Firstly, we need to preprocess the plate image using digital image processing. License plate image preprocessing includes grayscale, image enhancement, image smoothing and binarization. The thesis proposed a positioning method of texture feature based on license plate and prior knowledge. The method adopts level difference and vertical difference to find approximately location area. Then uses a priori knowledge to exclude the interference areas. In order to remove the fake license plate areas with counting number of pixels in transition of license plate. Finally, we can get accurate vehicle plate area.Secondly, we can reduce the computation and obtain good results by using contour extraction with Hough transform to tilt correction for the license plate image. Then remove the top and bottom license plate frame and rivets. This thesis applies the vertical projection and combining features of the character to segment the license plate characters. Normalized and tight rearrangement the segmentation characters.Thirdly, combination of genetic algorithm and neural network to recognize the license plate characters. Train the neural network using genetic algorithms to find optimal weights and threshold. The traditional genetic algorithm is improved by using real number coding method to enhance the network weights and threshold accuracy. At the same time, we use a variety of cross-way parallel crossbar, which is broadening the range of species and will help the search to the global optimal solution. Adaptive mutation rate on the one hand to ensure the diversity of species, on the other hand to make the algorithm rapid convergence to the global optimization. Experiments show that the algorithm greatly improves the learning efficiency and convergence speed.