摘要
复杂道路口车流量大,车牌采集图像存在遮挡,倾斜等现象,导致车牌检测准确率低、泛化性差。为解决上述问题,提出一种基于改进Yolov3车辆识别与Mask R-CNN车牌检测相融合的算法,通过两步分别检测车辆与车牌,最终构建出YOL3-RCNN多车牌识别模型。模型首先基于上采样算法将图像的大、中、小特征有机融合;然后采用形态聚类算法,自适应计算IOU阈值,提高车辆检出率与准确率;接着对检测出的车辆进行裁剪,SOFTNMS优化与仿射变换,解决车牌遮挡,倾斜与背景复杂的难题;最后基于ResNet50网络识别车牌字符与数字,并利用RPN进行回归校验。多组基线算法的仿真结果显示,在UA-DETRAC车辆检测数据集上,与传统车牌检测算法相比,YOL3-RCNN算法的准确率与召回率分别提高了7.7%与9.6%,且具有较高的车牌检测速率。综上所述,所构建的YOL3-RCNN多车牌识别模型具有较高的检测准确性、时效性与普适性。
In order to solve the problem of low accuracy and poor generalization of license plate detection,this paper proposes an algorithm based on improved Yolov3 vehicle recognition and Mask R-CNN license plate detection.By detecting the vehicle and license plate separately in two steps,a YOL3-RCNN multi license plate recognition model was ultimately constructed.The model first integrates the large,medium and small features of the image based on the up sampling algorithm,and uses the morphological clustering algorithm to adaptively calculate the IOU threshold to improve the vehicle detection rate and accuracy;Next,the detected vehicles are cropped,optimized with SOFTNMS and affine transformation to solve the problems of license plate occlusion,tilting,and complex background;Finally,the characters and digits of the license plate are recognized based on ResNet50 network,and the regression verification is carried out by using RPN.The simulation results show that the accuracy and recall of YOL3-RCNN algorithm are improved by 7.7% and 9.6%,respectively,compared with the traditional license plate detection algorithm on the UA-DETRAC vehicle detection dataset,and it has a higher license plate detection rate.To sum up,the YOL3-RCNN multiple license plate recognition model constructed in this paper has high detection accuracy,timeliness and universality.
作者
张晓辉
宋欧铭
郑维
于心俊
ZHANG Xiao-hui;SONG Ou-ming;ZHENG Wei;YU Xin-jun(School of Electrical Engineering,Henan University of Technology,Henan Zhengzhou,450001,China)
出处
《计算机仿真》
2024年第8期150-154,160,共6页
Computer Simulation
基金
河南省自然科学基金(202300410117)
中国博士后科学基金资助项目(2022M712382)
河南省科技厅重点科技攻关项目(192102210240)。
关键词
复杂道路环境
大流量路口
多车牌自动识别
Complex road environment
High traffic Intersections
Automatic recognition of multiple license plates
