The Improvement Research Of Vehicle License Plate Recognition System Based On Deep-learning

With the improvement of China’s economic strength and the development of comprehensive national strength,the number of vehicles released by the Ministry of Public Security reached 235 million units in November 2018,an increase of 12.0% year-on-year,including 1.01 million new energy vehicles,an increase of 83% year-on-year.The use of new energy vehicles and ordinary vehicles is increasing,and intelligent traffic identification becomes more and more difficult.Therefore,a good intelligent transportation system is a crucial step to solve this problem.How to identify different types of license plates in different environments efficiently and accurately is the key and difficult point.Only when we correctly identify the license plate can we better manage those license plates.License plate recognition technology is widely used in many aspects: the management of school entrance and exit vehicles and the management of vehicle information in and out of the gates of the community,and the speeding violation of driving at the crossroads.Through the mastery of all aspects of image technology,machine learning and deep learning,this paper improves the difficulty of the license plate recognition technology with the deep learning method.The following improvements have been made to new energy license plates and common license plates: license plate positioning module,correction module,character segmentation module and character recognition module.The recognition system uses a deep learning algorithm and framework,and the camera module on the hardware side has also been upgraded to meet the requirements of faster and more accurate identification.The main work of this paper is as follows:(1)In the license plate location module,a positioning algorithm with cascaded classifier as the coarse positioning and combined with the regression model is proposed.Firstly,the candidate regions are located,and the pseudo-license plates are removed by multi-level binarization and connected domain markers.The RANSANC algorithm and the CNN regression model are used to achieve accurate positioning.When the encounter is not possible,the CNN regression model is directly used for positioning.The method effectively overcomes factors such as illumination and tilt..In the off-line experiment,conventional license plates and new energy license plates collected in harsh environments such as different provinces and different illuminations were collected for testing.Among them,9000 conventional license plate positioning rates were 99.344%,and 964 new energy license plate positioning rates were 99.38%.(2)In the license plate character segmentation module,the ground color judgment and license plate correction of the license plate are first performed.Usually,the license plate background color is composed of 2 colors.The K-means algorithm obtains the main component color and the subordinate color.By extracting the main component color value and comparing the defined rule table,if it is blue,the 7-character sliding template is used,and if it is green,the color is selected.Character sliding template.Since the time complexity of the traditional transform correction algorithm is O(n^3)level,the speed of license plate correction processing is relatively slow.In order to solve the problem of adhesion under real-time performance,a fast license plate based on texture field is adopted.Text correction algorithm.The direction of the direction of the field of each sub-block is counted to find the direction in which the text is most dense as the direction of correction.When sliding the template,the optimal segmentation position of the license plate is the position where the sum of the probability values of all characters is the largest.In the off-line experiment,8919 conventional license plates were successfully segmented and corrected,the segmentation success rate was 99.04%,and the success rate of 957 new energy license plate segmentation was 99.06%.(3)In the step of license plate character recognition,the three algorithms of combination,SVM-Hog,SVM-CNN and CNN-GRU are used to study character recognition.At the same time,the old training sample library has been optimized and supplemented accordingly.The character recognition of 8834 conventional license plates and 948 new energy license plates after successful segmentation is CNN-GRU.The CNN-GRU algorithm has a recognition rate of 98.58% for conventional license plates and 98.95%,which is robust to similar character recognition.(4)In terms of overall system design,on the basis of the TensorFlow framework,the implementation of the platform is adopted by Py-Qt,and the industrial camera of the model is used to complete the real-time system realization of the modules such as license plate location,license plate segmentation and license plate character recognition.Through the actual measurement of 1512 conventional license plates and 213 new energy license plates,the conventional license plate successfully positioned 1,496 vehicles,the positioning success rate was 98.94%,and the new energy license plate successfully positioned 210 vehicles.The positioning success rate was 98.6%,which was earlier than the laboratory;on the basis of successful positioning,the conventional license plate segmentation success was 1478 vehicles,the segmentation success rate was 98.8%,the new energy vehicle segmentation success was 207,and the new energy license plate segmentation success rate was 98.57%.On the basis of correct segmentation,the characters of 1,460 conventional vehicles are fully recognized correctly,the character recognition rate reaches 98.79%,and the characters of 204 new energy vehicles are fully recognized correctly,and the character recognition accuracy rate reaches 98.55%.In summary,the overall measured recognition rate of conventional license plates is 96.56%,and the overall measured recognition rate of new energy license plates is 95.78%,which improves 13.41% of license plate recognition system compared with the previous versions.