Study On License Plate Recognition Based On Digital Image Processing And Machine Learning

With the rapid economic and social development, there come up a lot of problems in transportation. The number of various types of motor vehicles grows rapidly while the capacity of highway increases slowly and the management relatively lags behind which make transportation problems more serious than ever. Therefore, an advanced traffic administration and control system is urgently needed. Fortunately Intelligent Transportation Systems (ITS) which makes traffic more effective and secure at the same time can help. License Plate Recognition (LPR) is an important part of ITS. LPR is a kind of intelligent system that is able to recognize car license plates automatically. LPR can widely be applied for traffic flow observation, stolen cars tracing, parking management, illegal plates distinguishing, vehicle monitoring, etc.Aiming at distinguishing license plates, dividing characters and identifying separated characters under complicated background, this paper performs a deep study on key technologies used in license plates recognition.The following is the main research and development work of the thesis: Firstly, in order to achieve rapid recognition of license plates and to increase anti-disturbance capacity, this article studies AdaBoost technology, analyses the characteristics of license plates, and then it brings forward the AdaBoost algorithm based on jump-change count and variance of jump-change count. For increase of training effectiveness of sorting, construction of a strong classifier is of the utmost importance. To meet the demand for non-plate training samples during the construction process of strong sorter, a non-plate automatic generation system based on training model is constructed in this paper. It increases the diversity of training sample collection, which therefore is able to achieve the target of effective training of the strong sorter. This developed system can effectively obtain license plates in a complex environment. In various test set, it achieves consideratable performance.Secondly, by using OTSU a license plate image is converted to a binary image. Radon transform corrects slant of license plates and erases superfluous frames with future knowledge. The system divides the plate characters with vertical projection method. Then a license plate with low disturbance and basically regular is got.Finally, by analyzing the structure features of number and letter, this paper establishes a multi-level classification strategy which is based on a hole feature classification. The sub-class plate characters are sorted by BP ANN. Training of ANN based on genetic algorithm optimization avoids over-fit training samples set. Consequently, a high recognizing rate of the sorter is got.According to this study, the improved AdaBoost algorithm can better position license plates and has a relatively high ability of anti-disturbance. BP ANN has a better ability of generality.