Research And Implementation Of License Plate Recognition System In Unconstrained Scenes

Vehicle information management system is one of the important researches in the field of intelligent transportation.Vehicle license plate,as important identification of vehicles,has been an active research object in computer vision.With the increasing number of vehicles in recent years,the vehicle information management system based on automatic vehicle license plate detection and recognition has gradually began to be applied in traffic control,access management,vehicle query and other practical life.In this paper,the license plate recognition system is designed by analyzing the small target and interference of license plate recognition in the unconstrained environments.The validity of the license plate recognition system is verified by implementing several comparative experiments on the publicly available data sets.The main research work includes the following contents:(1)Research on the robust license plate detection algorithm in unconstrained scenes.A robust license plate detection method based on convolutional neural network is proposed to meet the challenges under different scenarios.Based on the classical object detection algorithm Faster R-CNN,this paper introduces the feature priors and contextual information of license plates into the network to achieve accurate license plate detection.First,in the feature extraction stage,the integrated feature map is established by combining convolution features of different resolutions,so as to improve the accuracy of multi-scale license plate classification and regression.Different from Faster R-CNN,this paper introduces multi-angle anchors and branch convolution structure in the RPN stage to generate proper candidate regions.Secondly,in order to deeply explore the relationship between license plate and its context,this paper proposes a context fusion network to further enhance the feature representation.Through a series of experiments,the detection algorithm proposed in this paper shows excellent performance of license plate detection on AOLP and SSIG data sets.(2)Research on the accurate license plate recognition algorithm in unconstrained scenes.By taking the whole string as recognition target in the license plate image,this paper proposes a vehicle license plate recognition algorithm based on LSTM and attention mechanism.The convolutional layer is designed for feature extraction,combining the long-short-term-memory network and the CTC sequence model,as well as the recognition of the seven-digit blue plate and eight-digit new energy license plate of unlimited length.Considering that the feature representation ability of single license plate character is insufficient,attention mechanism is further introduced to weighted representation of adjacent features of recurrent unit,which mapping to enhance character representation ability.In addition,this paper proposes a separator strategy to avoid the decodation problem caused by the same serial character in the CTC model.Compared with the current license plate recognition algorithm,the proposed method can avoid the error accumulation of segmentation and realize the license plate recognition with unlimited length from end to end.(3)Design and implementation of license plate recognition system under unconstrained scenes.Through analyzing the function and complexity of the license plate recognition system and the requirements of information display,the designed function module includes the input and display module,the license plate detection and recognition algorithm and the process control module.Through the practical test,the license plate recognition designed for the vehicle license plates in a variety of scenarios,which achieved 7% recognition improvement compared with the mainstream license plate recognition system Easy PR and Hyper LPR.