基于YOLOv5算法的满文不定长字元数据集制作方法研究
Research on the method of making Manchu indeterminate length character dataset based on YOLOv5
李昭仪 1于淼 1于晓鹏1
作者信息
- 1. 吉林师范大学数学与计算机学院,吉林 四平 136000
- 折叠
摘要
在进行满文识别时需要用到大量的满文数据,但目前还没有满文不定长字元数据集.本文提出一种基于YOLOv5的满文不定长字元数据集制作方法,用于后续的训练和研究.与传统切割方法相比,只需提供待检测图片即可.通过对基于YOLOv5的数据集制作流程的改进,去除原YOLOv5实验中对图像进行翻转和随机裁剪部分,并且将原YOLOv5的损失函数替换为EIoU,添加了注意力机制SE模块.实验结果表明:与原始的 YOLOv5网络相比,其精度和召回率分别提高到98.95%和98.83%,证明了算法的实用性和高效性.
Abstract
A large amount of Manchu data is needed in Manchu recognition,but there is currently no Manchu indeterminate length character dataset available.To solve this problem,a method of making Manchu indeterminate length character dataset based on YOLOv5 is proposed for subsequent training and research.Compared with traditional cutting methods,only the image to be detected is required.By improving the dataset making process based on YOLOv5,the parts of image flipping and random cropping in the original YOLOv5 experiment are removed.The loss function of the original YOLOv5 is replaced with EIoU,and the attention mechanism SE module is added.The experimental results show that compared with the original YOLOv5 network,its accuracy and recall rate are improved to 98.95%and 98.83%,respectively,which proves the practicability and high efficiency of the algorithm.
关键词
YOLOv5/EIoU/SE模块/数据集制作/目标检测/满文Key words
YOLOv5/EIoU/SE module/dataset making/target detection/Manchu引用本文复制引用
基金项目
&&(YDZJ202301ZYTS285)
出版年
2023
NETL
NSTL
维普
万方数据