摘要
针对藏式古建筑由于光照、复杂纹理等因素导致匹配误差而难以获取高分辨率深度图的问题,提出了自适应代价体聚合以及具有可靠注意力深度细化模块的深度学习网络。该深度学习网络采用从粗到精的策略,引入自注意力机制增强图像特征提取能力,并使用自适应代价体聚合减少像素匹配误差;通过深度细化提高深度图精度并减少累积误差,经过迭代得到高分辨率深度图。实验结果表明,在自建数据集上的重建图像较完整、纹理清晰,在DTU数据集上的准确度误差、完整度误差和综合误差分别为0.297、0.347、0.322 mm。基于多视图的三维重建可为研究和保护藏式古建筑提供有效帮助。
Aiming at the problem of difficulty in obtaining high-resolution depth maps for Xizang ancient architecture due to matching errors caused by illumination,complex texture and other factors,a deep learning network with adaptive cost volume aggregation and reliable attentional depth refinement module is proposed.This deep learning network adopts a strategy of from rough to fine,introduces self-attention mechanism to enhance image feature extraction ability,and uses adaptive cost volume aggregation to reduce pixel matching errors.By using depth refinement to improve depth map accuracy and reduce cumulative errors,high-resolution depth maps are obtained through iteration.The experimental results show that the reconstructed image on the self-built dataset is complete and the texture is clear,and the accuracy error,integrity error and comprehensive error on the DTU dataset are 0.297 mm,0.347 mm and 0.322 mm,respectively.3D reconstruction based on multi-view can provide effective help for the study and protection of Xizang ancient architecture.
作者
李建邦
杨晓波
许应成
罗骞
LI Jianbang;YANG Xiaobo;XU Yingcheng;LUO Qian(School of Information Engineering,Xizang Minzu University,Xianyang 712082,Shaanxi,China)
出处
《实验室研究与探索》
CAS
北大核心
2024年第8期29-34,46,共7页
Research and Exploration In Laboratory
基金
西藏民族大学科研项目(22MDY018)。
关键词
深度学习
三维重建
多视图
自适应方法
深度细化
藏式古建筑
deep learning
3D reconstruction
multi-view stereo
adaptive method
deep refinement
ancient Xizang architecture
