Research On License Plate Detection Method In Complex Scene Based On Deep Learning

In recent years,license plate detection algorithms have not only been used in fixed scenarios such as parking lots and high-speed toll stations,but have also been introduced into law enforcement tools such as drones,which will face more complex scenarios than before,such as severe weather,strong light,etc.These complex scenes bring huge challenges to license plate detection due to illumination,angle inclination,etc.Traditional detection algorithms are based on artificial features,which are not suitable for license plate detection in complex scenes.Therefore,this paper mainly studies the license plate detection algorithm based on deep learning.The specific research contents are as follows:(1)Firstly,although the existing open source license plate dataset CCPD has subsets of multiple scenes,the number of samples in each subset is not the same,so this paper constructs a new hybrid dataset LP-CCPD based on CCPD,which balances the number of samples in different scenes,and then manually annotates th e sample images for model training and testing of license plate detection tasks.(2)Secondly,several commonly used target detection algorithms are introduced in detail: Faster R-CNN,SSD,YOLOv3,YOLOv4,and their network structures and loss functions are analyzed.Then,a comparison experiment of license plate detection is carried out on the new mixed data set LP-CCPD,and the detection performance of these four algorithms is analyzed and compared.Through comparison,YOLOv4 is selected as the basic detection model for subsequent research on license plate detection methods in complex scenarios.(3)Then,an improved YOLOv4 license plate detection algorithm is proposed in view of the fact that the classical target detection algorithm cannot meet the high accuracy requirements of the license plate detection task in complex scenes.The main improvement measures are: first,in order to improve the detection accuracy of the algorithm and the detection effect of small target objects,improve the multi-scale feature detection network,and increase the original 3-scale feature prediction to4-scale feature prediction;second,in the backbone network Hollow convolution is introduced to expand the receptive field of extracting feature maps and promote the effective fusion of feature layers;third,the new data set is re-clustered through the clustering algorithm,and the size of the prior frame is optimized and adjusted.The improved model is trained and the license plate detection experiment is carried out.The analysis of the experimental results shows that the improved model has an ideal detection effect in the license plate detection task in complex scenes,but the detection speed is lower than the original YOLOv4 algorithm.(4)Finally,to solve the problem of slow detection speed of the improved algorithm,a lightweight license plate detection algorithm with YOLOv4 is proposed.Based on the improved algorithm in the previous chapter,the network model is lightened,a lightweight convolutional neural network is in troduced as the new backbone network,and the standard convolution in the enhanced feature extraction network is replaced by a depth-separable convolution.