Research On The Key Techniques Of License Plate Recognition

License plate recognition system is at the core of the Intelligent Transport, can be seen its application everywhere in our daily life. Because of widely applicable, not only has brought considerable economic value to the product, but also the study of theory and technology systems are more depth and refinement along with the people increasing attention. This paper explored and analyzed the traditional algorithm of the core steps in the license plate recognition system, on this basis, puts forward some effective improvements. Core Steps of license plate recognition system mainly includes a list of work: image preprocessing, license plate location, characters segmentation and recognition of license plate number.In the positioning license plate area, in order to obtain a clear picture with small storage space, the need for preprocessing to the original license plate images collected from the device. According to the interference factors of collected images background have a greater impact on the license plate location, the paper studies a positioning based on texture and color characteristics of the license plate. In the process of coarse location based on texture features, through horizontal black and white jump scanning and the edge of the vertical density characteristics analyzing determines candidate areas. Using the three component values of HSV space to identify the correct license plate area. This algorithm can exclude fake license plate with a similar texture features and also avoid error positioning because of the body with the same color of license plate.In the step of characters segmentation, considering the low accuracy of the previous single license plate characters segmentation algorithm with common problems, this paper studies the improved 8 connected component labeling algorithm combined with vertical projection algorithm to achieve self-adaptive characters segmentation. This paper improved labeling algorithm that most of the characters can be applied with one scanning. This algorithm overcomes the problem of needing a memory of big volume to memorize connecting relationships and a great deal of operations to unite labels. The application of this algorithm compared to traditional segmentation algorithm not only improved on the speed, but also can effectively solve characters of adhesion and unconnected Chinese.In the step of characters recognition, According to the strokes of Chinese characters are various and complex, resulting in Chinese characters recognition accuracy rate is much lower than the alphanumeric. In this paper put forward feature extraction for Chinese characters based on the PCA, for alphanumeric based on the gird algorithm. The feature vector of the classification was inputted into the BP neural network for training, by adding momentum term and adaptive learning rate to improve it. In order to improve the recognition accuracy, according to the different positions of the plate corresponding to the different classes of character. We design and train three classifiers, they are the Chinese characters classifier、the capital letter classifier、the number and capital letter mixed classifier. Through three kinds of classifiers to classify, until the completion of the training target output the recognition results, thus completing the license plate number recognition steps. After experimental analysis can be obtained, the algorithm of this paper can effectively solve the problem of low Chinese characters recognition rate, the time of recognition can also meet the real-time need.