Research On Low Resolution License Plate Recognition Algorithm

In the intelligent transportation system,shooting a vehicle from a surveillance video is often only because the captured vehicle is far away,and can only capture low-resolution license plates,while the Chinese character structure of the license plate at low resolution is seriously degraded,which greatly affects The license plate recognition accuracy rate.Therefore,how to accurately identify the license plate number of the captured vehicle and provide the vehicle identity information for the management of intelligent traffic information has become an urgent problem to be solved.Since the license plate resolution of the traveling vehicle captured in the surveillance video is very low,and the external environment such as image tilt and character blur caused by the shooting angle is disturbed,it is difficult to accurately position and divide the license plate.To this end,this paper uses the modified detection network Faster R-CNN to locate the license plate,which improves the accuracy of license plate location.Aiming at the problem of unrecognizable caused by the structural degradation of low-resolution license plate Chinese characters,this paper proposes a two-level low-resolution license plate recognition method based on cyclic neural network and twinning network.The first level uses the character without division to start the license plate.Identification,firstly,the deep neural network VGG16 is used to extract the more robust depth features of the license plate,and then the obtained depth features are converted into feature sequences.Finally,the feature sequences are sent to the cyclic neural network(bidirectional LSTM)to obtain the initial recognition results of the license plate.In the second stage,by using the feature that the license plate has a large gap at the focus,the Chinese character part that is prone to recognition error is positioned and divided,and only the Chinese character and the adjacent letter area are retained,and the divided Chinese characters are re-identified.Firstly,through the initial recognition of the license plate,the candidate Chinese characters of Top1 and Top2 corresponding to Chinese characters are obtained,and then the Siamese twin network is constructed.The Chinese characters and candidate templates to be recognized are simultaneously input into the twinning network,and the similarity of the image pairs is output.Finally,the similarity is obtained.The template corresponding to the output is the result of the secondary recognition of the Chinese character.The experimental results show that the proposed algorithm has a recognition accuracy of over 98%for the entire license plate,especially the Chinese character part,and the effect is obviously superior to other commercial license plate recognition systems.