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

Automatic license plate recognition(ALPR)can detect license plates and recognize characters from images.As one of the important technology of the intelligent traffic management,ALPR has been widely used in daily life.But the traditional license plate recognition relies on the artificial designed features,causing it is difficult to detect and segment license plates accurately in complex scenes with unstable illumination,variable resolutions and angles of images.The deep learning can adaptively extract abundant features of images,which provides a new thought for the license plate recognition in complex scenes.In order to prove the feasibility and practicality of applying deep learning to the license plate recognition in complex scenes,this thesis does these work as follows:(1)As there is not a unified dataset for license plate recognition in complex scenes.This thesis captures a large number of license plate images in multiple scenes,angles and weather.Then,for getting more samples,this thesis does some data augmentation.Finally,the license plate images are manually labeled to make the final data set for the experiment.(2)Because the traditional technology of license plate detection cannot detect license plates effectively from complex scenes.By comparison of many kinds of target detection algorithms,this thesis proposes a network named LPD-net based on YOLOv2 object detector for license plate detection.For designing the prior boxes of license plates,this thesis applies the clustering algorithm to the data set.For enhancing the detection performance of small targets,this thesis uses multi-scale detection strategy.Finally,this thesis combines the characteristics of different convolution layers to enhance the ability of feature expression of the network.The results of experiments show that,compared with the traditional method,LPD-net has better performance in complex scenes and reaches the speed of real-time.(3)In view of the deficiency of traditional method of ALPR,this thesis proposes an end-toend license plate recognition network named LPR-net based on Faster R-CNN.Firstly,the Region Proposal Network(RPN)is used as the license plate detection network to get areas of license plates.Secondly,the bidirectional recurrent neural networks(BRNN)is used to identify characters of license plates.Compared with multi-step methods of ALPR,the endto-end method can unify the training process of the license plate detection network and the character recognition network.These two networks share the basic convolutional layers,which reduces the number of parameters and improves the speed.The end-to-end license plate recognition network of this thesis outperforms the traditional methods in datasets,which proves the feasibility of deep learning in dealing with license plate recognition in complex scenes.