Research On License Plate Location And Recognition Algorithm Based On Convolution Neural Network

With the progress of science and technology,intelligent transportation system plays a more and more important role in urban management,but in practical application,the existing recognition system is still insufficient in efficiency,accuracy and other aspects,so the accurate recognition of vehicle recognition in smart city still has research significance.Based on the research of the existing recognition algorithm,this paper puts forward some improvement suggestions as follows.First of all,this thesis improves the object detection method YOLOv3(you only look once version 3)and proposes the YOLOv3.1 algorithm to locate the license plate.The activation function of YOLOv3 algorithm is Leaky Relu(Leaky Rectified Linear Unit),the backbone network is residual network.The sum of Location information loss function,Class loss function and Confidence loss function is used as the loss function of YOLOv3,and the optimization is made on this basis:(1)The original Leaky Relu activation function is replaced by the Mish activation function(A Self Regularized Non-Monotonic Neural Activation Function),and Mish activation function avoids the saturation caused by touching the vertex,and the loss of effective information of neural network,and enhances the robustness and generalization ability of the network;(2)Dense Net is used to replace the original residual network in the backbone network,and the characteristic of Dense Net is to obtain more effective information by dense connection;(3)The loss function is updated to a loss function that fuses the Focal loss function with the Complet Intersection over Union.The fusion method is to replace the Intersection over Union of YOLOv3 with the Focal loss function as Confidence loss function,which can effectively improve the detection effect of the network for difficult samples.And use the CIo U loss function to replace the Sum of Squared Error(SSE)of YOLOv3 as location information loss function to increase the sensitivity of the model to the location information of the border and the overlapping way of the prediction box and the actual labeled border,which can improve the detection accuracy and accelerate the convergence speed of the model.Through the comparison of the follow-up experiments,the thesis finds that the YOLOv3.1 license plate detection model improves the mean average precision(m AP)by 4.07% compared with the original YOLOv3,which indicates that in the aspect of license plate detection,the improvement of the model brings a certain improvement to the performance of license plate detection,and completes the accurate license plate target detection.Then,this thesis applies the target detection network structure CSPResNeXt proposed by Wangchenyao et al.to license plate recognition.A license plate recognition model based on CSPResNeXt is proposed.CSPResNeXt is mainly used in image detection and classification,which has excellent performance in target detection,but the parameter setting based on target detection can not be directly applied to license plate recognition.This thesis adjusts it as follows:(1)By analyzing the length width ratio of domestic license plate,it is found that the existing input layer resolution is not suitable for license plate recognition,which causes stretch or compression deformation to the image.After calculation,the average resolution of license plate is 72×273,so the resolution of 224×224 is changed to 72×273 in the input layer,which can effectively retain the license plate information;(2)the default parameter of target detection network in the output layer is1000.Detect domestic 7 bit license plate the output layer parameter 1000 is changed to 476.Each line of vector recognizes one character as the output.If the parameter is not modified,it increases the calculation of the network structure.So CSPResNeXt network structure adjusts the output layer parameter 1000 to 476,which can improve the parameter utilization and reduce the calculation of the network structure;(3)The analysis of CSPResNeXt network structure shows that it is not sensitive to small scale targets,but RFB can imitate human visual perception.Therefore,CSPResNeXt network structure add RFB module at the sampling place of 8 times and 16 times of feature map,so that license plate information in complex scenes can be better identified.After the above three adjustments,the CSPResNeXt system is formed for license plate image recognition.The results show that the accuracy of license plate recognition model based on CSPResNeXt is99.10%,which is higher than Rese Net,ResNeXt and Dense Net,and has strong learning ability.And CSPResNeXt network model is smaller than Rese Net,ResNeXt and Dense Net,even half the size of ResNeXt,and the operation efficiency is higher.Compared with the experimental test of Diaoyanhua et al.on the model of license plate recognition based on Open CV,the experimental results show that the recognition accuracy of the license plate recognition model based on CSPResNeXt is improved by 7.3%.Finally,this thesis proposes a new license plate recognition system named YLCR tandem system,by cascading the YOLOv3.1 location model and the license plate recognition model of CSPResNeXt.First,YLCR tandem system inputs the license plate image into YOLOv3.1 for positioning,and then inputs the image after positioning to the license plate recognition model of CSPResNeXt for recognition.The experimental results show that the final recognition accuracy of YLCR tandem system is 94.13%,and the average recognition time of each license plate image is0.096 seconds,which is less than 2 seconds required by GPU and fully meets the equipment requirements.Moreover,the YLCR tandem system has a wider scope of application and more robustness.It can accurately locate and recognize the license plate in some complex scenes,such as too strong or too dark light,occlusion,and tilt angle.Compared with it,the recognition effect of traditional recognition system in some complex scenes is not ideal.Finally,this thesis compares the YLCR tandem system with the open source traditional license plate recognition system Easy PR and the license plate recognition system Hyper LPR based on deep learning.The experimental results show that the accuracy of Easy PR is only 85.2%,the accuracy of Hyper LPR is 92.5%,and the accuracy of YLCR tandem system is 94.13%.It can be seen that the YLCR tandem system has a high accuracy.