基于Faster-RCNN和先验知识的车架VIN码识别方法  

FRAME VIN CODE RECOGNITION METHOD BASED ON FASTER-RCNNAND PRIOR KNOWLEDGE

作　　者：赵珣 张新峰 边浩南 Zhao Xun;Zhang Xinfeng;Bian Haonan(Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China)

机构地区：[1]北京工业大学信息学部,北京100124

出　　处：《计算机应用与软件》2024年第1期177-182,共6页Computer Applications and Software

摘　　要：为了提高车检所工作效率,同时克服对VIN(Vehicle Identification Number)这类长字符串识别准确率低的难题,基于现有深度卷积神经网络的模型,提出以Faster R-CNN为主干网络,并结合先验知识的车架VIN识别模型。根据车架号图像特点,选择Faster R-CNN进行字符级定位和识别的方案。针对长字符识别容易漏字符的现象,使用被遗漏位置的前后字符坐标来定位缺失字符。使用inception网络对补漏得出的字符区域进行识别。灵活使用先验知识使得该方法比只使用Faster R-CNN识别车架号的准确率提高了31.7百分点,识别率达到了64.77%,这也高于当前主流OCR模型在长度超过15位的文本上的准确率。In order to increase the working efficiency of vehicle inspection office and overcome the problem of low accuracy of long character string recognition,we proposed a VIN(Vehicle Identification Number)recognition model based on the existing deep convolutional neural network model,which combines Faster R-CNN as the backbone network with the prior knowledge of VIN image.According to the characteristic of VIN image,Faster R-CNN was selected for character level positioning and recognition.In order to solve the problem of missing characters in long character recognition,we used the coordinates of the characters before and after the missing position to locate the missing characters.We used the inception network to recognize the character regions obtained from the omission.By using prior knowledge flexibly,the accuracy of our method is 31.7 percentage points higher than the model of using Faster R-CNN only,the recognition rate reaches 64%,and that is also higher than the accuracy of prevailing OCR model in the recognition of character string that longer than 15.

关 键 词：OCR Faster RCNN 先验知识 长字符串 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]