Research On Algorithm Of Multi-license Plate Recognition Under Complex Background And Its Software Implementation

In recent years, with the development of "smarter city" concept, and rising figures of domestic car ownership, how to collect and process information of vehicle scientifically and efficiently, make it into the operation of the whole "smarter city" system has been an important topic in the field of scientific research and business. As the first part of the intelligent system, the license plate recognition algorithm and the software system are the most important. As an application in the field of machine vision, license plate recognition system can be applied in the treatment of illegal traffic, vehicle tracking, intelligent parking lot, etc. It has a very important significance to improve the efficiency of urban management.By studying and comparing the domestic and foreign related algorithms. It is found that the domestic research and application of multi-license plate recognition system still lags behind foreign related field, it has significant research value and business prospects. Considering the plate under a complex background with interference of rotation and deformation, a kind of multi-license plate recognition system is designed, detailed as follows:(1)By comparing the performance of four region moment descriptors quantitative by the ROC curve(receiver operating characteristic curve), and select two kinds of the regional moment descriptors that optimal performance. Then form a new feature vector using SFS(Sequential Forward Selection). Finally, the feature is classified and predicted by SVM. License plates can be located efficiently and accurately by the new way, even the target has a change of location, attitude or scale.(2)By studying the method of license plate character segmentation based on image projection, it is found that only the license plate under ideal condition can be segmented by the standard license plate image projection method. In view of the situation of adhesive and polluted characters, a dynamic peak and valley method based on image projection method is proposed. The problem of license plate character segmentation under non ideal conditions can be solved effectively.(3)In the character recognition process, by comparing the performance of standard BP neural network and improved BP neural network algorithm with different number of neurons in the hidden layer, and different growth rate of adaptive learning rate. The optimal structure of the BP neural network is determined and the performance of character recognition reaches the best state.(4)Complete the requirement analysis and system design. Visual Studio 2010, MFC and MS Access were used to build a three layers system structure based on C/S, and accomplished the overall development of the license plate recognition system, which is used a large amount of image processing knowledge, the auto recognition of multi-license plate system has been implemented successfully. The result demonstrates that the improved multi-license plate recognition system is accuracy and practicability.(5)Finally, on the basis of static image recognition, the method of license plate recognition based on video stream data is discussed. The extraction and optimization methods of single frame image are presented. The experiment proved that the operation speed of the system is improved.