Research On Tilt License Plate Correction And Recognition Based On Deep Learning

With the rapid development of artificial intelligence,how to extract license plate information more effectively has become an important task in traffic management.Therefore,finding an algorithm with strong robustness and high recongnition accuracy based on the image recognition technology is an important improvement direction of license plate detection algorithm.This technology mainly processes the extracted features then classifies them.Generally speaking,the adaptive effect of traditional detection methods is not good,the detection speed is hardly to meet the actual requirements,and the development of detection algorithms is difficult,besides,the development period is long.In practical applications,the requirement of more effective extraction of features has seriously restricted the development of traditional methods.The use of deep learning technology can solve these problems very well.This technology is a data-dirven mothod,the relevant features are extracted through trainng a large number of image to generate the deep learning model.This thesis mainly includes the following parts:Firstly,the writing structure of this paper is to start with the overall framework and use CNN model based on deep learning to detect and locate license plate corners.As far as the detection model,the model trained by the various state datasets constructed in this paper is compared with the traditional edge location algorithm,color location algorithm and edge color location algorithm,which verifies the superiority of the model.Secondly,based on the projective geometry theory,this thesis studies the problem of tilt image correction in license plate correction.The research shows that most of the tilted license plates are converted into standard images by pixel interpolation,and some of the interpolation regions are blurred.Therefore,the perspective tilt license plate is converted into affine tilt license plate by reducing the interpolation ratio.The number of real pixels in the original image is increased,and then the IoU value of the license plate area is increased by window transformation.Finally,the image pixel translation operation is used to correct the orthogonal license plate,which greatly improves the clarity of the image pixel and the IoU value of the license plate location.It lays a good foundation for the subsequent recognition work.Compared with the ROI selection algorithm,the experimental results show the superiority of the correction effect in this paper.Thirdly,the recognition model trained by CRNN network is used to recognize the characters in the detected license plate area.Based on the traditional recognition technology,this recognition technology derives an end-to-end method.It does not need to segment the characters on the license plate.It takes the license plate as an image,sorts the above characters,and analyses and studies them as a whole.Finally,the results are given.This paper illustrates the overall framework of license plate recognition based on CRNN model.After correcting the license plate,CRNN model can be used to identify the license plate.Then,the detailed description of the training method,the structure of CRNN network model and the implementation process are presented.Finally,the experimental results are obtained and analyzed.Experiments show that CRNN model is an effective method for image sequence recognition.Through the verification of many experimental results,we can see that the traditional recognition technology has a lot of drawbacks,which relies on character segmentation heavily.Once the segmentation is inaccurate,serious errors will occur,which greatly reduces the robustness of the algorithm.In contrast,this paper combines various functions of the deep learning model in three aspects: license plate location,license plate correction,license plate recognition,and further optimize the performance,the robustness and accuracy of the algorithm are improved gradually,demonstrates the concept of intelligent city and intelligent transportation.