| 摘 要: 针对文献中手写体甲骨文的检测工作数据集空缺、目标占比低等问题,构建了首个文献中手写体甲骨文的检测数据集,并基于 YOLOv8提出了一种基于双层路由注意力的检测方法YOLO-SA(YOLO with Stratified Attention)。该方法利用滑窗裁剪技术提升目标占比,引入双层路由注意力模块增强模型对甲骨文有效信息的提取,并采用SIoU损失函数替代原损失函数,提升小目标检测的准确度。实验结果表明,YOLO-SA在自建数据集上的精确率达到90.2%,召回率达到92.3%,相较基线方法分别提升了17.9%和17.7%,证明自建数据集的实用性和所提方法的有效性。 |
| 关键词: 文献图像 甲骨文检测 数据集构建 YOLOv8 注意力机制 损失函数 |
|
中图分类号: TP391.41
文献标识码: A
|
|
| Research on Handwritten Oracle Bone Inscriptions Detection Method in Literature Based on Bi-Level Routing Attention |
|
SHI Zhan1, LI Jian1, YANG Jun1, TANG Pei1, WANG Yongshan1, WANG Shuai2
|
(1.School of Electronic Information and Artificial Intelligence, Shaanxi University of Science and Technology, Xi’an 710021, China;
2.School of History and Culture, Shaanxi Normal University, Xi’an 710062, China)
221612161@sust.edu.com; lijianjsj@sust.edu.cn; 211612101@sust.edu.cn; 221611058@sust.edu.cn; lswx0925@gmail.com; 447561195@qq.com
|
| Abstract: To address issues such as the lack of datasets and the small relative size of targets in handwritten oracle bone inscriptions detection in historical literature, this study constructs the first detection dataset for handwritten oracle bone inscriptions in literature. Based on YOLOv8, a detection method named YOLO-SA is proposed, incorporating b-i level routing attention. This approach first employs sliding window cropping to increase the relative size of targets. Subsequently, a b-i level routing attention module is introduced to enhance the model’s ability to extract effective features from oracle bone inscriptions.
Finally, the original loss function is replaced with the SIoU loss function to improve detection accuracy for small targets.Experimental results demonstrate that YOLO-SA achieves a precision of 90.2% and a recall of 92.3% on the custom dataset,representing improvements of 17.9% and 17.7% over baseline methods, respectively. This validates the practicality of the sel-f built dataset and the effectiveness of the proposed method. |
| Keywords: literature images oracle bone detection dataset construction YOLOv8 attention mechanism loss function |
