Research On End-to-end License Plate Detection And Recognition Algorithm Based On Feature Fusion

With the rapid development of big data,artificial intelligence,internet of things and 5G technology,as well as the rapid development of China's urbanization,the construction of smart cities has become a trend.Intelligent transportation plays a very important role in smart cities.It can not only alleviate urban traffic congestion,but also improve the use efficiency of traffic roads,save resources and improve the ecological environment.In addition,the traffic monitoring system also plays a role in protecting the city.The system can help the public security organs quickly track suspicious vehicles and find criminal suspects who escape,thus making the city we live in safer.As the license plate number is equivalent to the ID Card of a car,the license plate detection and recognition algorithm technology plays a very important role in the construction of intelligent transportation system.The traditional license plate detection and license plate character recognition algorithms can achieve good recognition effect in some simple scenes,but they are difficult to achieve ideal detection and recognition effect in complex scenes such as insufficient light,tilt,blur and bad weather.Many researchers usually put the license plate detection and recognition process is divided into two steps,the license plate detection,license plate recognition or into the license plate detection,license plate character segmentation,license plate character recognition three steps.But in front of license plate character recognition and the license plate detection and license plate character segmentation steps are highly correlated,the current step of license plate detection algorithm to predict the license plate boundary frame error,will lead to the loss of license plate characters,or cause behind the step of license plate character segmentation error,so that eventually led to the back of the license plate character recognition step predict error of license plate number.In order to solve the defects of the traditional license plate detection and recognition algorithm,this paper will target detection algorithm based on convolution neural network is introduced into the license plate detection and recognition,design a kind of based on feature fusion of end-to-end network model license plate detection and recognition.The network model structure is composed of license plate character extraction,feature fusion network,multi-scale characteristics of the network,license plate location identifier and license plate character recognizer composed of five parts.In this network model,the feature maps in different layersare fused to preserve the semantic information of license plate position on shallow feature maps,which is aimed at further improving the recognition accuracy of license plate position and characters.Compared with the license plate detection and recognition method which is divided into several steps,the network model can be trained from end to end.When a license plate picture is input,the boundary box of license plate position and license plate number can be output at the same time,so as to avoid the problem of intermediate error accumulation and error amplification caused by traditional license plate detection and recognition methods.As the network model extracts the network convolution layer through the shared basic features,it reduces the network computing parameters,thus significantly improving the speed of license plate detection and recognition.In this paper,a license plate data set is made for the training and testing of the network model,in which license plate images are collected under different complex scenes.Through the experimental comparison of different algorithms,it is proved that the fusion feature method can effectively improve the accuracy of license plate location and license plate character recognition.