The Research And Implementation Of License Plate Automatic Recognition Technology

In this paper, we mainly discussed the design and implementation of vehicle license plate recognition system. According to the character of real imaging environment of license plate, we integrative used the technology of image analysis and artificial neural network, and brought forward a complete solution.Following work has been finished in this dissertation:1. We introduce the present situation and commonly used methods about license plate location and character recognition at home and abroad. After comparing these methods, we bring forward a new LPR method based on improved feature of rough grid and BP neural network learning.2. On the license plate location, we introduce the traditional image processing technology, for example, gray conversion, gray extension, gray equipoise, edge-finding methods etc. We use the both PREWITT operator and CANNY operator to detect the intensity binary edges, and returns a binary image of the same size, with 1's where the function finds edges and O's elsewhere. We also use the subtraction of gray values from adjacent pixels and median filter to weaken the white noise on the background of license plate. Based on the distribution of white leap change points, we use the rectangle matching method to gain the approximate location area of license plate. On certain condition, We use an adaptive method to search the edges up and down the rows, and compare long with broad so as to search right-and-left edges. At last, we expand the license plate image to realize the exact location and cutting of license plate.3. For the preprocessing of license plate image, we introduce a fast algorithm to get the binary image threshold, and use RADON transformation for incline revise and geometry location adjusting of license plate. This method overcomes the shortcoming of traditional HOUGH transformation, and has a high efficiency. The HOUGH method is always interfered or fooled by median point and noise with ease. We introduce a method to unify the background color of license plate image, and bring forward four rules to wipe off frame and rivet. Through reverse rotation, we can remove burrs. At last, we use the erode and dilate operation to unify the thickness of license plate character image and prepare for the follow-up character segmentation and recognition.4. On the character segmentation algorithm, we use the connected charts to remove noise with mark rectification algorithm, which reduces the time of image scan and makes the characters on the license plate clearer. We also use the high efficiency ZHU's integration license plate character segmentation method to segment character. This method can availably resolve the problem of how to segment the conglutinationor incline characters correctly. We use vertical projection and trough algorithm to scan image fast and to do uprightness segmentation. While coming up against mistake segmentation or if failure because of incline image, we use the dripping water algorithm to do again along the contour curve to do our best to reduce segmentation failure rate.5. We choose the feature of rough grid as the feature of character recognition, and directly input the improved unified character primitive feature to BP neural network classifier to recognize the license plate character. To improve the recognition rate of the character recognition system of vehicle license plate, we analyze the result of test, and put forward a series of measures. For example, we design a careful neural network classifier to distill the detail features of character which are analogical and easy to confuse. Aimed at the phenomena of stroke conglutination and character excursion found in Chinese characters, based on standard stylebook, for some province characters which are complex in structure and compact in strokes, we suitably add representative stroke conglutination stylebook and representative excursion stylebook to the network training stylebook of Chinese character. Experiment results show that the method based on BP neural network classifier, especially the integration of rough classifier with fine classifier can maximally improve the robust performance of character network, upgrade the ability of anti-jamming, tolerance and correctness for the whole vehicle license plate recognition system.6. In this paper, we also discuss the design method of BP network framework, provide the number of input layer nerve cell and output layer nerve cell, especially the number of hidden layer and hidden layer nerve cell, and the choosing of activation function and the setup of parameter of all neural networks. We also describe the process of training and learning. Based on the summarization of experience of predecessors, we bring forward a new method to judge the constringency degree of network synthetically, and then to adjust learning parameter and momentum gene, to control the time to stop training rationally, which uses the change about error energy mean observed in the process of serial training and the difference between the real value and the target value of network output nerve cell. We design four neural networks (Chinese character network, letter network, letter & numeral network and numeral network) to classify character, which improve the efficiency and recognition rate.7. Based on the research of theory and combined with our scientific research project, we have developed an automatic recognition system modeling for vehicle license plate and implemented the correspondence algorithms using Matlab 6.5 and Visual C++ 6.0 programming tools.