| The container handling capacity ranking of Chinese ports has been continuously improved,The real-time performance and security of the original operating system need to be further improved.After the vehicle enters the gate,it is necessary to realize a comprehensive monitoring of the transportation of goods.And it could master and view all movement tracks of the vehicle in time,so as to realize the comprehensive perception,automatic identification and unmanned operation of the vehicle entering the gate.In order to realize the automatic collection of vehicle license plate information and the construction of gate unmanned monitoring system.This paper will study and implement vehicle license plate recognition based on computer vision according to the samples of Container terminal.For specific application technologies,this paper mainly introduces the following research contents and innovation points from three basic steps of target area positioning,character segmentation,image classification and convolutional recurrent neural network(CRNN):(1)Aiming at the difficulty of license plate location in port area,this paper studies the target location method based on image processing and the target location algorithm based on deep learning.In the method of image processing and positioning,this paper proposes a method of color space separation based on H and S channels,which are sensitive to the color of the license plate itself,and the corresponding relationship between the background of the license plate and the characters.This paper studied the location method of Adaboost+SVM.Because the results of the screening model are closely related to the samples,this paper proposed that two SVM candidate region judgment models were trained by using the different size of the coincidence degree with the target region when making the training positive samples,so as to obtain more accurate license plate region when locating the license plate as much as possible.An improved SSD model is proposed in the deep learning positioning model algorithm,and the network structure of the feature layer used to generate the default box in the SSD framework is modified to better adapt to license plate samples of different size and resolution and different types.Finally,this paper adopts the improved algorithm of Adaboost+SVM and SSD based on color space separation to locate the license plate region,with an accuracy of 98.7% and a single average positioning time of 0.32 s.(2)To solve the problem of character segmentation failure caused by damage,blur and small character spacing in character area of truck license plate in port area,this paper proposes fixed binary method of color background statistics and MSER&OSTU binary method for character segmentation.Firstly,the gray scale of the license plate area is stretched to improve the contrast,and then the characters are separated from the background by binarization method.Then the upper and lower rivets of the license plate are removed by the number of black and white jump of the binary image,and then a single character rectangle is obtained by using the connected domain segmentation.At last,the four boundaries of the rectangle are adjusted by using the rectangle width,height and spacing to get an accurate single character subgraph.In this paper,the success rate of license plate character segmentation is94.3%,and the average segmentation time of single license plate is 0.28 seconds.(3)Aiming at the problem of multiple license plate types and diverse character composition,this paper studied the classification model of ANN and CNN neural network for license plate recognition.Through analysis and comparison,the single-character recognition rate of Lenet5 network model reached 99.5%,and the average single-character recognition time was 0.018 seconds,with better effect.In order to meet the requirements of multiple types of license plate character recognition,such as standard and internal license plate system,this paper USES Lenet5 network to train three types of character classifiers according to the number of classification tags and design a reasonable character recognition process.Finally,the correct recognition rate of license plate character is 96.7% and the average recognition time of license plate character is 0.22 seconds(4)In order to further improve the license plate recognition accuracy,the traditional segmentation part of license plate recognition method is unable to correctly identify problems,this paper studies the CRNN network direct recognition algorithm,to locate the license plate area,successfully extracted features using convolution layer,circulation layer prediction probability distribution characteristics results and tag sequences,transcription layer for the tag sequences with the highest probability.The test results show that CRNN can successfully identify 37.1% of the unrecognized samples in the traditional segmentation algorithm,and improve the overall license plate recognition accuracy by 3.5%.The license plate recognition data sets used in this paper are all from the actual scene of container terminals.After locating the original sample license plate area with multi-algorithm combination strategy,the traditional segmentation recognition and CRNN direct recognition of the license plate area are carried out respectively in the way of multi-threading,and the final output result is determined by comparing the recognition result and confidence degree of the two.The accuracy of the identification results in this paper is 93.6%,and the average identification time of the single sample is 1.19 s.In practice,it can meet the requirements of accuracy and real-time... |
PDF Full Text Request