FEC-PVT:基于PVT架构的甲骨钻凿图像分割网络

摘要：

由于长时间埋藏于地下和风化腐蚀,造成甲骨片破损和甲骨钻凿边界模糊不易分辨,给甲骨钻凿分割带来极大挑战.从甲骨数据库及著录书中系统收集并标注甲骨钻凿图像.基于该数据集,提出一种以Transformer为编码器的甲骨钻凿分割网络FEC-PVT(feature extraction and connection pyramid vision transformer),首先,FEC-PVT利用FE-C和FE-D模块分别补充低层和高层特征,以获取细节和全局特征;其次,FCOM模块用交叉注意力让不同层特征交互,获取有效细节;最后,FFDM模块逐层解码并整合多层次特征,提升解码精度,避免特征丢失.实验验证,所提FEC-PVT优于其他的方法，与次优的DuAT方法相比,IoU提高5.18%.

Due to long-term burial in the ground and weathering and corrosion, the oracle bone fragment is damaged and the boundary of the oracle bone drill is blurred and difficult to distinguish, which brings great challenges to the division of the oracle bone dill and dhiseI. This study Systomattcally collects and annotates oradle bone drill and chisel  images from oracle bone databases and cataloged books, Based on this dataset, this paper proposes a feature extraction and connection pyramid vision transformer(FEC-PVT) with Transformer as encoder. FEC-PVT uses FE-C and FE-D modules to supplement low-level and higb-level features, respectively, to obtain detailed and global features. Secondly, the FCOM module uses cross-attention to allow different layer features to interact and obtain effective details, Finally, the FFDM module decodes layer by layer and in-tegrates multr-level features to imprtove the decoding aceuracy and avoid feature loss. Experimental verifitcation shows that the FEC-PVT proposed in this paper outperforms other methods. Compared with the suboptimal DuAT method, its loU inereasesby 5.18%.

作者: 

刘国奇，李文格，茹琳媛，宋黎明，刘杰，韩燕彪

Zhu Guifen, Du Jiejie, Luo Chenshi, Chen Runan, Liang Pengfei, Wei Yaxin

机构地区：

河南师范大学 a.计算机与信息工程学院；b.甲骨智能计算实验室；c.历史文化学院

引用本文：

刘国奇，李文格，茹琳媛等。FEC-PVT:基于PVT架构的甲骨钻凿图像分割网络[J].河南师范大学学报(自然科学版),2026,54(1):8-16.(Liu Guoqi,Li Wenge, Ru Linyuan,et al.FEC-PVT:An oracle bone drilling image segmentation network based on PVT architecture[J].Journal of Henan Normal University(Natural Science Edition),2026,54(1):8-16.DOI:10.16366/j.cnki.1000-2367.2024.12.10.0003.)

基金：

国家自然科学基金；河南省高校科技创新团队；河南师范大学杰出青年科学基金；河南师范大学优秀科技创新团队

关键词：

image segmentation; oracle bone drilling; pyramid vision transformer; convolutional neural networks

分类号：

TP391

FEC-PVT:基于PVT架构的甲骨钻凿图像分割网络.pdf