Research On License Plate Recognition Algorithm In Complex Scene Based On Deep Learning

With the rapid development of economy,the total number of vehicles in China is increasing day by day,License plate recognition has become a key technology for traffic control transportation management.Based on traditional license plate recognition algorithms,it can be applied to simple scenarios such as residential areas and parking lots;however,in the all-weather complex working environment,the vehicle license plate image will be blurred,damaged,slanted and so on,the traditional algorithm anti-jamming ability is poor,the model is not universal,the algorithm location and recognition effect is not ideal.Because of the limitation of the traditional method,this paper presents a license plate recognition algorithm based on deep learning.First of all,this paper proposes a method based on improved u-net semantic segmentation model to extract the license plate area.Semantic segmentation is a method based on pixel-level classification,which is often used in some high-precision scenarios such as medical treatment,compared with the target detection model,the semantic segmentation network model is simpler,with fewer network parameters,and a small amount of data set can achieve better results.Experimental verification shows that semantic segmentation model based on improved u-net by adding attention mechanism can further increase the weight of the region of interest and improve the detection accuracy.Through the data screening in each scene of the public data set,a total of 3000 pictures containing vehicles were selected and sent to the segmentation model for training.The test results show that the license plate location algorithm based on the semantic segmentation model has Mean intersection-parallel ratio.Then,through the analysis of the license plate area,this paper realizes end-to-end character recognition based on improved Le Net’s convolutional neural network.The improvements to the network structure are as follows:(1)to replace one 5 * 5convolution in Le Net with two 3 * 3 convolution,which has the same ability of feature extraction,the latter has stronger ability of expression,fewer parameters and less computation;(2)adding a global average pooling layer to the connections of the convolution layer and the full connection layer of the network,so that the network can accept input of any size;(3)Change its head to seven parallel fully connected layers,sharing the previous convolutional layer.Through this improvement,the license plate can be recognized end-to-end without the need for character segmentation processing.The test results show that the accuracy of the improved Le Net on the license plate is more higher.Finally,we concatenate the license plate location algorithm based on the Improved U-Net algorithm for license plate location and the license plate character recognition algorithm based on the improved Le Net.The whole model is tested and the result shows that the accuracy of the final license plate recognition is up to99.15%.And the whole model can meet the requirements of real-time detection and recognition.