License Plate Recognition Based On Unit Coupled Pulse Coupled Neural Network

Vehicle License Plate Recognition (LPR) system is an important part in Intelligent Transportation System (ITS). ITS began to reflect its values as vehicles increased in China. LPR has been used extensively in unattended garage, campus security, community security, traffic monitoring of restricted areas and so on. LPR do release pressure of persons on duty. Therefore, the researches on LPR become hot spots in these days.The LPR consists of three steps,1).Plate Localization(PL);2).Character Segmentation(CS), and3).Character Recognition(CR). These three steps are serial and the later one impacted by the former one, any robust step can remedy others. Studying these three steps together is needed in order to improve the whole performance. Some algorithms of LPR are reviewed in this thesis and mainly three preposed algorithms are introduced in detail.i. A Plate Localization method is proposed. One channel of RGB of an input car image is processed by Unit-Linking PCNN which output a fired image. Then, areas containing the features of plate are located from the fired image. The ratio of plate width to lenth is also included. At last, the area of interesting will be tilt corrected. The output of PL algorithm is a binary image.ii. The projection algorithm for plate character segmentation is improved using Unit-Linking PCNN. The model of PCNN using for hole filling is the same as which is used in above mentioned PL method. After filling hole in plate characters, connected components on the middle line of plate image are found. The size of this plate characters are normalized by analyzing the connected components. At last, the boundary coordinate values of each characters are determined together with projecting and analyzing the connected components. The improved segmentation method make segmention rate and recognition rate both higher because the characters segmented from one plate are normalized.iii. It is proposed that characters can be recognized by using PCNN fired image sequences together with the number of holes counted by PCNN as features. The difference between the features of unknown characters and that of template characters is measured by voting. The same Unit-Linking PCNN is used to thin the characters and fired the thinning image. Hole numbers are counted by PCNN as well. Different weight are given to each features. By voting for unknown character, recognition result is made finding the highest voted template. Three sub-algorithms consist the LPR methods. In sub-algorithm PL, CS and CR, Unit-Linking PCNN is all used. The structure of PCNN is all the same, just initiate and decision conditions are changed. No more parameters are setted which was did in other PCNN application. Unit-Linking PCNN is the core for PL and CR sub-algorithms and in CS, it improved the original algorithm. All above just verify that Unit-Linking PCNN has potentials and advantages in concrete applications of image processing.