Research On The Key Technologies Of Automatic License Plate Recognition

Abstract:Along with economic development and continuous automobiles increasing, there is an urgent need to improve the automation degree of vehicle monitor and manage in the transport system. License Plate Recognition (LPR) is one of very effective method and it has high values on theory research and practical application.For the purpose of monitoring highway speeding vehicles, algorithms about license plate localization and recognition are proposed in this thesis which mainly studies the license plate images shot at highway speeding monitoring points. Because images’resolution is quite little, and the license plate in the image is also little, the research is very difficult. Main research results and innovations are as follows:A license plate localization algorithm is proposed, based on region gradual narrowing, with high adaptability and accuracy. The algorithm mainly contains distinguishing between day and night, seeking vehicle left and right borderlines, and seeking license plate left, right, top and bottom borderlines. In the algorithm, distinguish between day and night images of license plate with gray-scale histogram smooth curve. Seek vehicle left and right borderlines, removing noise with longitudinal difference binary image. Seek license plate top and bottom borderlines, calculating transverse difference image, and removing influence of ventilation cover. Then the plate left and right boundaries are determined by the differential image horizontally. The test results of practical application show this algorithm can effectively solve the daytime and evening highway speeding vehicles smaller image plate positioning problems, and it can get a higher rate of license plate location positioning.A plate area binarization and a characters’segmentation algorithm are presented in this thesis.A license plate binarization method based on the the mean variance experience threshold is proposed, which can get good binarization effect for license plate region. For characters’ segmentation, we calculate segment attributes with longitudinal projection smooth extended curve, separate a long segment to three segments or two segments, unite two short segments, delete dots and noise, adjust segment attributes to fit the character segmentation request, and then extract, adjust the characters left, right, top and bottom borderlines. The experimental results show that this method is more efficient, more stable, relative to the commonly used algorithm.A character normalization algorithm is proposed. Assigning object serial numbers to character binary image pixels, the small areas is removed as noises. Then surrounding blanks of character are excluded to shrink character borderlines. The large deformation problem of character "1" after zooming is solved by adjusting aspect ratio. Finally we get the normalized character images by shrinking Chinese character to24×32matrix, shrinking letter or number to16×16matrix. Normalized same characters have a high degree of coincidence, excluding noise and interference, conducive to improve the recognition rate with same size.BP neural network training and recognition algorithms of Chinese characters, letters and numbers are presented in this thesis. BP neural network is adopt three-layer structure, including input layer, hidden layer and output layer.A lot of characters are trained according to the standard BP neural network steps including activate computing, error calculation, connection weights adjustment and threshold adjustment. Then the characters are recognized according to activate computing, target vector calculating, and character serial number calculating.The experimental results show that this method makes the BP neural network training speed faster, and recognition rate even higher.