Research On License Plate Location And Recognition Based On Improved CNN

In recent years,China’s per capita income level continues to improve,cars began to enter thousands of households,at the same time,the increase in the number of vehicles has also brought great pressure to traffic management.Therefore,the demand for intelligent management of traffic system is very urgent.As an important part of intelligent traffic system,the location and recognition of license plate is very key to its efficiency and accuracy.Due to the different scenes of license plate shooting,the license plate itself has some problems,such as tilt,wear,blur,etc.,usually the traditional license plate location system is not good for the license plate location.On the other hand,traditional license plate recognition requires character segmentation of license plate,which is too complicated and has low recognition accuracy.In view of the above problems,this paper improves the existing methods,and the main work is as follows:(1)To solve the problem of inaccurate license plate location,this paper uses image pretreatment,morphological transformation,edge detection and other methods,combined with the license plate size information,color information to locate the license plate.The specific method is as follows: preprocessing the image with Gaussian function,finding the image edge through binarization and edge detection,opening and closing the edge to connect the edge into a whole,and finally screening the rectangular frame through the length-width ratio information and color information in HSV space to obtain the license plate location result.The results show that the accuracy and recall rate of this method are 7.11% and 5.02% higher than those based only on color and morphology,and 4.78% and 3.43% higher than those based only on edge and morphology.(2)In view of the problem that license plate recognition requires character segmentation and thus affects the recognition accuracy,this paper adds BN layer,introduces PRe LU activation function,and further introduces STN module,BLSTM network and CTC transcription layer based on VGG16 model.The improved STCRNNS model realizes the end-to-end license plate recognition without the need of license plate segmentation.Compared with Res Net,Goog Le Net,VGG16,and CRNN,the recognition accuracy of this model in MIX data sets is improved by 2.98%,1.81%,3.91%,and 0.86%,respectively.