| As an important part of intelligent transportation system(ITS),license plate recognition technology in the natural scenes can provide efficient and practical management methods for intelligent transportation.However,for the traditional license plate recognition system,the algorithm has limited anti-interference ability and low robustness.License plates can't be accurately located and recognized due to uneven illumination,plate tilt and fouling and license plate images with complex background in natural scenes.This paper conducts research on these issues using convolutional neural network(CNN).Firstly,this paper proposes a target detection method based on CNN to locate the license plate.Besides,the network topology is optimized to satisfy the practical application of license plate detection,which significantly reduces the network parameters.8000 license plate images in natural scenes are manually labelled for network training.Test results significantly demonstrate that license plate location based on the optimized CNN has high accuracy and recall.In order to improve the generalization ability and robustness of the detection network,the license plate images which can't be located successfully are further trained to enhance the network generalization ability.Secondly,the optimized license plate detection network is used to crop the license plate images.Furthermore,data augmentation is implemented by generative adversarial network(GAN)to enrich the training images,which can effectively avoid the case that the training set is too small lead to the network model over-fitted and improve the network generalization ability.The main superiority of this scheme is that it is segmentation free and the cropped license plate images are directly input into the recognition network for feature extraction and classification.Then,this paper studys the Residual Network(ResNet)and Dense Convolutional Network(DenseNet)topology from the lens of high order recurrent neural network(RNN)and proposes a simple,efficient and modular network topology.Based on this topology structure,the license plate recognition scheme based on deep CNN is proposed.The recognition network introduces group convolution as ResNeXt to enhance the leaning capacity of each micro-block,and at the same time implements the dense connection fr-om layer to layer based on DenseNet topology.The new network structure inherits the advantages of ResNet and DenseNet,besides,it can efficiently reuse features and flexibly explore new features.Extensive experiments verify the effectiveness and robustness of proposed scheme,and the recognition accuracy achieves 99.34%in complex natural scenes.Finally,the optimized detection network and the well-trained recognition network are concatenated for the overall license plate recognition test.The overall test shows that the license plate image localization accuracy can achieve 99.98%and the recognition accuracy can achieve99.34%in complex natural scenes.Besides,the average recognition time of the whole system is less than 0.1 second per image using GPU,which can fully satisfy the practical application. |
