The Study Of License Plate Image Segmentation And Recognition Algorithm

With the development of science and technology, People's lives are more and more intelligent and automated. There are more and more vehicles, which increase the pressure on the highway charges and district management. While the intelligent traffic surveillance (ITS) can change this situation, improve the efficiency of traffic surveillance, make it scientific and standardized. License plate recognition (LPR) technique is the kernel part of ITS. License plate recognition consists of three modules in general:plate location, character segmentation and character recognition. License plate recognition is one of the research subjects in which the technology of computer vision and recognition applied. The purpose of transporting systems is to facilitate the automatic management. Therefore, LPR is the key technology to achieve the demand of ITS.The thesis studies mainly the latter two modules of the license plate recognition system--character segmentation and character recognition, and the recognition results are obtained by experiments.First, there must be pre-processing for the vehicle license images before segmentation of the characters. In this phase, it has studied the gray image conversion, tilt correction, image binarization and so on. Tilt correction is mainly used an algorithm based on linear-rotation-projection to detect the tilt angles caused by wrong directions in which we got these license plate images, then the tilted license image is correct. A global threshold got by OTSU method is used in Binarization to achieve a good binarization effect. Then the border of the upper, lower, left and right and the split point in the license image are removed with a black-white-jumping algorithm. After that the improved vertical projection method is used to segment image, and the character images after segmentation are normalized to get a unified size of character image. Finally, the final recognized result of these character images are got with a three-layer BP neural network. The network is divided into there classifiers, such as Chinese characters, letters, numbers and letters. The fine-mesh characteristics of a large number of samples are extracted to train the BP network. Then we can get the recognizing result by sending the characteristics of fine-mesh of the recognizing character images into the BP neural network, and recognize these character images.