Research And Application Of Mobile License Plate Recognition Based On Deep Learning

License plate recognition algorithm based on deep learning is much better than traditional algorithms in robustness and accuracy.However,due to its high computational requirements,it is difficult for license plate recognition algorithm based on deep learning to achieve both accuracy and speed.In this paper,the license plate detection and recognition in the video collected in the open environment are studied to meet the accuracy and real-time.In this paper,by analyzing the data characteristics of the existing public license plate data set,a certain number of virtual license plates are generated on the basis of CCPD data set combined with the license plate images collected in the field.Through license plate image enhancement,tilt correction and other technologies,the license plate detection and recognition basic data set is established.In the aspect of license plate detection and location,this paper analyzes the advantages and disadvantages of several common license plate location algorithms at home and abroad.Aiming at the problems of insufficient feature extraction and low feature utilization in the license plate detection task of Yolo lightweight network,this paper finally designs a mobile license plate detection algorithm yolov4 tiny LPD based on yolov4 tiny,and proposes a bottom-up multi-scale fusion Combine the low-level information to enrich the feature level of the network and improve the feature utilization.The mobile license plate collected in open environment will be more difficult to recognize because of the sensitive problems such as illumination and angle.Based on the investigation and study of license plate character recognition algorithms at home and abroad,combined with real-time recognition based on video,convolution neural network is applied to license plate recognition algorithm to realize the end-to-end recognition of license plate characters.According to the specific technical requirements of license plate recognition,convolution neural network is directly applied to the pixel information of the whole license plate image to train and learn the pattern of the whole license plate and get the recognition results of license plate without character segmentation,so as to cope with the complex and changing environment.After a large number of test data verification,the accuracy of the proposed license plate detection algorithm model TinyLPD in the test set reaches 99.96%,with an average time of 0.009s;the accuracy of CNNLPR in the sample test set of this paper can reach 99.01%,with an average time of 0.017 s.Experimental results show that the proposed license plate detection model TinyLPD and license plate character recognition network model CNNLPR have high accuracy,real-time performance and strong robustness in natural environment.