Research On Multi-scene License Plate Recognition Technology Based On Deep Learning

License plate recognition is one of the key technologies of vehicle recognition and an important part of intelligent traffic.Although license plate recognition has been mature in hd static traffic scenes,it still faces challenges in broader scenes.Therefore,how to quickly and accurately obtain license plate number information in the complex and changing scene has great practical value and significance.The traditional license plate recognition is based on the artificial design features,which is not robust enough in complex scenes such as unstable illumination and oblique camera Angle,so it is difficult to realize the high-precision recognition.In recent years,the deep learning method has made great progress in various tasks of computer vision and provided new ideas for the research of multi-scene license plate recognition.Therefore,based on deep learning,this paper designed a reasonable neural network structure and strategy for license plate recognition in multiple scenes,so as to improve the recognition accuracy and efficiency and meet the needs of different scenes.The main work and research achievements of this paper are summarized as follows:tilted license plate recognition based on spatial transformation network.Aiming at the problem that the traditional license plate recognition method cannot be effectively generalized to the tilted license plate scene,this paper use the spatial transformation network to correct the tilted license plate.The network can be trained end-to-end with the recognition network,and no additional labeling information is needed.The experimental results show that the proposed method can effectively correct the tilted license plate,thus improving the recognition accuracy,and only increases the calculation amount by 1.52%.License plate recognition based on composite data.For deep learning algorithm needs a large number of labeled data driven to train model,but artificial data collecting plates is insufficient and uneven distributed problem,this paper designed the license plate image synthesis method to the training set data are targeted,balanced distribution of the training set to build all kinds of data,at the same time of high precision accuracy,saves the consumption of time required to collect and annotation data and human.License plate recognition based on lightweight separable convolution network.For the license recognizer based on the large network performance is good,but the large amount of storage space and computing,difficulty in the mobile terminal or car chip application problems,this paper introduces the depth separable convolution structure(depthwise separable convolution)to realize the lightweight network design,and designed a wide convolution module to quickly get the context information.The experimental results show that this method can reduce the computational amount of the model greatly while maintaining a high identification accuracy,and can be effectively deployed in the terminal where computing resources are relatively scarce.license plate recognition based on multi-task neural network.For the multi-task license plate recognition scenario,this paper designs a multi-branch recognition model of Shared basic network,each branch is used to realize different functions.The network can not only realize the license plate number recognition,but also complete the license plate color and license plate type discrimination,while expanding the network function,but also effectively save computing resources.It is difficult to realize license plate recognition in complex scene based on traditional license plate recognition method.Based on the method of deep learning,this paper designs the corresponding recognition algorithm for different license plate recognition application scenarios,which meets the practical requirements of license plate recognition in different scenarios.Moreover,the network designed in this paper has the advantages of small computation,high robustness and multi-task.