改进UNet++的瓷器文物显微气泡分割  

Microscopic bubbles segmentation of porcelain relics based on improved UNet++

作　　者：刘阳洋 耿国华[1,2] 刘鑫达 李展[1,2] 路正涵 LIU Yangyang;GENG Guohua;LIU Xinda;LI Zhan;LU Zhenghan(National and Local Joint Engineering Research Center for Cultural Heritage Digitization,Northwest University,Xi’an 710127,China;Institute of Visualization Technology,Northwest University,Xi’an 710127,China)

机构地区：[1]西北大学文化遗产数字化国家地方联合工程研究中心,陕西西安710127 [2]西北大学可视化技术研究所,陕西西安710127

出　　处：《西北大学学报(自然科学版)》2025年第1期129-138,共10页Journal of Northwest University(Natural Science Edition)

基　　金：国家自然科学基金(62271393);陕西省教育厅一般项目(19JK0842);虚拟现实技术与系统全国重点实验室(北京航空航天大学)开放课题基金(VRLAB2024C02)。

摘　　要：对瓷器文物显微气泡的分割,可以更加清晰地观察瓷器表面微观气泡的形态、数量以及分布规律,进而辅助文物专家进行瓷器碎片分类和文物鉴定等工作。但瓷器显微图像中气泡复杂多变,大小及分布不均匀,现有图像分割方法难以适应瓷器显微气泡特征。因此,该文提出一种基于卷积激活单元的网络AGUNet++,该网络重新设计密集跳跃连接,节点间采用Z字形连接方式,充分提取图像语义特征,防止信息丢失;同时,在卷积单元的密集跳跃连接处,结合注意力门控模块Attention Gate提出卷积激活单元CAU,增强与瓷器文物显微气泡分割任务相关的气泡区域学习,抑制不相关的区域;在训练过程中对每一层子网络的输出采用深度监督和交叉熵损失,有效增强瓷器文物显微气泡特征提取能力,细化分割结果。该方法在SD-saliency-900以及PRMI数据集上的实验结果表明,与经典图像分割网络相比,AGUNet++在MIoU、Precision、Recall和F1分数中均有一定的提升,表现出更好的分割效果。The segmentation of microscopic bubbles of porcelain relics can provide a clearer observation of the morphology,quantity,and distribution of micro bubbles on the surface of porcelain,which is of great significance in assisting relic experts in classifying porcelain cultural relic fragments and identifying porcelain cultural relics.However,the bubbles in porcelain microscopic images are complex and varied,with uneven size and distribution.Existing image segmentation methods are difficult to adapt to the characteristics of porcelain microscopic bubbles.Therefore,a network named AGUNet++based on convolution attention unit is proposed.This network utilizes a zigzag connection approach between nodes to fully extract image semantic features and prevent information loss.Meanwhile,a convolution attention unit is introduced by combining the dense skip connection of the convolution unit with the attention gate.The CAU enhances the learning of bubble regions relevant to the task of microscopic bubble segmentation in porcelain artifacts while suppressing irrelevant regions.Deep supervision and cross entropy loss are applied to the output of each sub network layer during the training process,which effectively enhance the ability to extract microscopic bubble features in porcelain artifacts and refine the segmentation results.The experimental results of this method on the SD-saliency-900 and PRMI demonstrate that AGUNet++exhibits certain improvements in MIoU,Precision,Recall,and F1-score,showing better segmentation performance compared to classical image segmentation networks.

关 键 词：瓷器文物显微图像 显微图像分割 UNet++ 注意力门 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]