Research On End-to-end License Plate Detection And Recognition Algorithm Based On Lightweight CNN

In recent years,with the development of intelligent transportation system,demand for license plate recognition is increasing which puts forward higher requirements for license plate recognition system’s task accuracy and operation efficiency.This thesis takes the license plate in the natural scene as the research object,and aims to design an end-to-end license plate detection and recognition system based on convolutional neural network.The main work and contributions of this thesis are itemized as follows:(1)After the summary of the design philosophy of lightweight CNN,SGNet,a lightweight backbone network is proposed.SGNet focuses on the lightweight convolutional modules and the idea of feature reuse to exchange low-cost operation for high-performance benefits.Therefore,in the subsequent comparative experiments,SGnet has achieved competitive accuracy and performance advantages in four different classification data sets;(2)A complete license plate detection algorithm based on convolutional neural network,SGDet is designed.The algorithm adopts the detection strategy of single-stage,single-scale and anchor-free strategies.SGDet has made important improvements in three aspects: sample allocation mode,boundary box regression form and multi-stage feature fusion and enhancement module.Benefited from the simple and efficient detection algorithm,SGDet can achieve the running speed of 170 fps under the condition of large-resolution input,and also can achieve 77% detection precision on the CCPD data set;(3)A lightweight license plate recognition model based on CNN and integrated with attention mechanism which improves feature extraction is designed.This part focuses on the design of SPP＿ROI Align module,which is able to extract the multi-scale features.SPP＿ROI Align module can provide rich and effective feature information for the recognition network and improve the recognition accuracy by 5%;(4)The holistic end-to-end license plate detection and recognition algorithm,CLPNet is finally synthesized that includes backbone network,detection network branch and recognition network branch.Further a dynamic task weight adjustment strategy is designed to improve the training effect.CLPNet can achieve 55.5% recognition accuracy and 160 fps running speed on CCPD data set.The comprehensive performance has great advantages over the current mainstream license plate recognition algorithms.On the self built data set LP350,CLPNet verified its outstanding generalization ability with 91.4% detection accuracy and 83.7% recognition accuracy.