Research On License Plate Recognition Technology Based On Support Vector Machine And Deep Learning

License plate recognition is a computer vision technology that uses the knowledge of digital image processing and pattern recognition to complete the identification of a license plate.The license plate is equivalent to the ID card of the car,and it is supervised by a unique identifier bound to the vehicle.It makes traffic management more efficient and convenient.Based on the investigation of traditional license plate recognition technology,this thesis makes a complete and detailed discussion on the four major technologies of license plate location,effective license plate judgment,character segmentation and character recognition with purpose of solving the problems existing in the traditional license plate recognition algorithm.Thus,the main research contents are presented as follows:1.This thesis discusses the limitations of single feature location algorithm,such as edge or color feature,and introduces the method of text feature location,which can make up for the shortages of edge and color feature location.Based on the consideration of the comprehensive application scenario and algorithm efficiency,the author adopts the methods of combining edge,color and text features to conduct research on license plate locate localization.2.When locating the license plate,it is impossible to filter out all the interference in the real scene regardless of the combination of edge,color and text features or other license plate location methods because some interference images similar to the license plate may be produced.In order to select effective license plates from candidate images,support vector machine model in machine learning field is used to classify and train the candidate images to determine whether they are license plates or not,and then the final license plate image is obtained.3.Because the license plate in China with special Chinese characters,requires that the character segmentation algorithm needs to be consisted with Chinese.After the investigation of relevant data,the author uses the combination of contour and prior knowledge to realize license plate character segmentation,and locates the position of Chinese characters by translating special characters,so as to make up for the poor support of traditional segmentation algorithm for Chinese characters.At the same time,all the segmented license plate characters are processed into 20×20 pixel size,the picture of the pixel size can display the license plate characters normally and clearly,which reduces the computational load of the network without affecting the recognition rate of license plate characters.4.In this thesis,the author applied related technology of deep learning to his research,and proposed an improved convolutional neural network LeNet-5 model for license plate character recognition.Through TensorFlow deep learning framework and TensorBoard visualization tool,the training accuracy and loss function comparison curves of the original convolutional neural network LeNet-5 model are drawn,and the parameters of the original convolutional neural network LeNet-5 model are modified.At last,the test results of the data show that the final improved convolutional neural network LeNet-5 model has higher character recognition rate,better feature classification effect and fewer training iteration steps compared with the traditional artificial neural network model.Based on the above research,the main algorithm's recognition effect is: in the license plate location part,using the edge,color,text features combined with the SVM model of the license plate location method,the license plate positioning rate in the conventional environment can reach 96.8%,in relative In the complex environment,the license plate location rate can reach more than 89.9%,and the FScore index of the trained SVM model reaches 96.84%.In the license plate character segmentation part,the license segment character segmentation method using character contour combined with prior knowledge is used,and the effective segmentation rate of Chinese characters is up to 96.6%,the effective division rate of English letters and numbers is 97.8%;in the license plate character recognition part,using the improved convolutional neural network LeNet-5 model,the final Chinese recognition rate is 99.2%,and the recognition rate of English letters and numbers is up to 99.6%.