Three-stage Natural Scene License Plate Recognition Based On CenterNet And ResNet

Aiming at the problem that most of the existing license plate recognition models are based on specific occasions and have poor generalization ability,this paper establishes a three-stage license plate recognition model composed of vehicle detection,license plate detection and license plate recognition based on Center Net,Res Net.At each stage,the existing model is optimized so that the model can achieve high-precision recognition of license plates in natural scenes.The main idea of three-stage algorithm is as follows :(1)The first phase of the model uses Center Net to build a vehicle detection model.The model speed is improved by simplifying the Hourglass-104 structure in the commonly used vehicle detection network Center Net.At the same time,the dilated convolutional network is used instead of the traditional convolutional neural network,and an output branch is added to the nonlinear activation function of the residual block to optimize the accuracy of the network.The average accuracy of the final model on the PKU dataset is achieved 58.9%,which is greatly improved compared with other classical vehicle detection algorithms.(2)The second stage of the model is improved on license plate detection model which based on Res Net.By adding WPOD-NET network in front of the model to correct the license plate distortion caused by the shooting problem,and using parallel output in the prediction stage,the detection accuracy and speed are improved.The average accuracy of the proposed model on PKU dataset reaches 95.26%,far exceeding other classical license plate detection models.(3)In the third stage of the model,a license plate recognition model is constructed using a CNN network that inserts the Res Net module.Firstly,Bi LSTM is selected as the sequence feature extraction module for the domestic license plate character rule,and then the parallel input Attention mechanism network which can save training time is used to construct the prediction module.Finally,the accuracy of the combined model on the PKU dataset is achieved 97.5%,which can effectively meet the high-precision requirements of license plate recognition.The three-stage license plate recognition model based on target detection and target recognition in natural scenes constructed in this paper has achieved the accuracy 98.1% on the PKU dataset.Compared with the existing license plate recognition model in natural scenes,the accuracy is greatly improved.