基于推荐区域的民国纸币冠字号码检测  

Detection of Crown Number of Republic of China Banknotes Based on Recommended Regions

作　　者：杜欣 王笑梅[1] 沈成龙 王佳婧 DU Xin;WANG Xiao-mei;SHEN Cheng-long;WANG Jia-jing(College of Information,Electrical and Mechanical Engineering,Shanghai Normal University,Shanghai 201418,China)

机构地区：[1]上海师范大学信息与机电工程学院,上海201418

出　　处：《计算机仿真》2023年第9期519-524,共6页Computer Simulation

摘　　要：针对传统的文本检测和一般的目标检测算法无法精准定位民国纸币冠字号码的问题,提出了一种融合异常值检测、聚类和目标检测的序列检测算法。首先定位推荐区域,利用YOLOv4算法确定字符分割块的位置;然后处理推荐区域,将字符分割块转换为数据点,利用局部相关积分(LOCI)算法检测并剔除异常数据点,采用算法Mean Shift算法聚合分割区域,提高定位精度;最终实现民国纸币冠字号码的检测。实验表明,上述组合算法检测的平均精度均值(mAP)达到99.6%,平均检测一张图像耗时0.83s,验证了算法的高效性。具有重要实际的应用价值。Aiming at the problem that the traditional text detection target detection algorithm cannot locate the character recognition of the banknotes in the Republic of China,a sequence detection algorithm that combines outlier detection,field of view and target detection,is proposed.First,locate the positioning algorithm,use the YOLOv4 algorithm to determine the position of the coded segmentation block;Then process the segmentation area,convert the coded segmentation block into data points,use the local correlation integral(LOCI)algorithm to detect and remove abnormal data points,use the algorithm to averagely translate the segmented area,and improve the positioning accuracy;Finally realize the detection of the serial number of the banknotes of the Republic of China.Experiments have proved that the average average accuracy(mAP)of the combined detection reaches 99.6%,and the average detection of an image lasts for 0.83 seconds,which verifies the effectiveness of the algorithm and has important practical application value.

关 键 词：民国纸币 推荐区域检测 目标检测 均值漂移 异常检测 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]