Research On Deep Faster-RCNN For License Plate Recognition

As a core technology of intelligent transportation system,license plate recognition has been widely applied to specific scenes such as high-speed toll gate and highway bayonet.How-ever,under the environment of weak illumination,low resolution,large tilt angle and large-scale background interference,the performance of the existing license plate recognition sys-tem still cannot meet the requirements of the application.The Research on deep Faster-RCNN for License Plate Recognition algorithm has important application value.License plate recognition includes two parts:license plate area detection and license plate number identification.The main contributions of this article are as follows:First,the license plate region detection is regarded as the two kind of target detection problem in the region of the license plate and the background area,and a license plate area detection algorithm based on Faster-RCNN is proposed.In terms of convolutional feature extraction,experiments were performed with ZF-Net,VGG-16,and ResNet-101,respectively.The results show that ResNet-101 has the highest accuracy.In the subsequent classification stage,the Soft-NMS method was used to suppress the non-maximum value,and the detection rate of the target was improved.Finally,according to the open data set Caltech Datasets,the comparison experiments were carried out using the algorithm of this paper,the license plate detection algorithm of the document 43 and the license plate detection algorithm of the document 44,which proved the effectiveness and advancement of the proposed method.Through the ex-periments of the bayonet license plate data set Normal Datasets,the validity of the proposed algorithm for scenes with fixed scenes,viewpoints and illumination is verified.In addi-tion,for the license plate dataset Hust Datasets,based on the dynamic changes of scenes,viewpoints and lighting,a comparison experiment was conducted using the algorithm of this paper,the license plate detection algorithm of the document 44 and the SVM license plate detection algorithm in the open license plate recognition system EasyPR.The accuracy of the algorithm in this paper is 97.2%,which is superior to other algorithms.Secondly,the license plate number recognition is regarded as a structured multiobjective detection prob-lem of license plate characters and backgrounds,and a license plate number recognition algorithm based on Faster-RCNN is proposed.The license plate number is obtained through the structural characteristics of license plate characters that are equally spaced on the same line and Soft-NMS processing.This algorithm circumvents the precise location and charac-ter segmentation and improves the accuracy and speed of license plate recognition.In the homemade Hust-Plate Datasets data set,the algorithm of this paper,the license plate number recognition algorithm of Tian LPRS in the literature 44 and the ANN license plate number recognition algorithm in EasyPR were compared.The accuracy of the algorithm in this pa-per reached 94.4%,which is superior to other algorithms.Finally,combining the license plate area detection and license plate number recognition algorithm,a license plate recog-nition algorithm based on depth Faster-RCNN is proposed.For the public data set Caltech Datasets,a comparative experiment was conducted using the algorithm of this paper and the deep learning license plate recognition algorithm of the literature 45,and the effectiveness of the algorithm was verified.Through the experiments of Normal Datasets,the validity of the proposed algorithm for scenes with fixed scenes,viewpoints and illumination is verified.In addition,on the self-made Hust Datasets data set,using the algorithm of this paper,license plate recognition algorithm in Tian LPRS and license plate recognition algorithm based on SVM and ANN in EasyPR were compared,the accuracy rate of this algorithm is 91.8%,which is superior to other algorithms.