| 刘韵婷,高宇,戴佳霖,谭明晓.融合ECA注意力层和轻量正则化的多视图三维重建[J].电子测量与仪器学报,2024,38(7):179-186 |
| 融合ECA注意力层和轻量正则化的多视图三维重建 |
| Multi-view 3D reconstruction combining ECA attention layer and lightweight regularization |
| |
| DOI： |
| 中文关键词: 三维重建 多视图 深度学习 ECA注意力 |
| 英文关键词:3D reconstruction multi-view deep learning ECA attention |
| 基金项目:辽宁省自然科学基金项目(2022-KF-14-02)、国家重点研发计划（2017YFC0821001-2）、辽宁省教育厅面上项目（LJKMZ20220617）资助 |
|
|
| Author | Institution |
| Liu Yunting | School of Automation and Electrical Engineering, Shenyang University of Technology, Shenyang 110159, China |
| GaoYu | School of Automation and Electrical Engineering, Shenyang University of Technology, Shenyang 110159, China |
| Dai Jialin | School of Automation and Electrical Engineering, Shenyang University of Technology, Shenyang 110159, China |
| Tan Mingxiao | School of Automation and Electrical Engineering, Shenyang University of Technology, Shenyang 110159, China |
|
| 摘要点击次数: 42 |
| 全文下载次数: 304 |
| 中文摘要: |
| 为了有效解决多视图三维重建中的边缘缺失、网络内存消耗严重、重建精度低的问题，对基于深度学习的多视图三维重建网络的特征提取、正则化网络、损失函数、优化器等进行研究。首先，使用融合ECA注意力层的特征提取网络，提高网络对通道特征信息的关注；然后，在门控循环单元模块中加入卷积层，改进的门控循环单元组合成GC正则化网络，采用此网络对代价体进行正则化，降低网络的计算量；最后，使用SmoothL1损失函数和Adam优化器，提高模型训练后期的收敛精度，优化模型的损失和参数。在DTU公开数据集上训练和测试，提出的融合注意力机制和轻量正则化的多视图三维重建网络（EGF-MVSNet）相比于经典的MVSNet网络完整性上提高了22.1，模型总体评分提高了11.5%。能够实现物体的重建，显著提高重建结果的质量，降低网络对内存的消耗。 |
| 英文摘要: |
| In recent years, multi-view 3D reconstruction technology based on deep learning has become one of the research hotspots in the field of machine vision and is applied in many fields. However, the 3D reconstruction technology still has problems such as edge missing, serious network memory consumption and low reconstruction accuracy. In this paper, based on the existing problems of 3D reconstruction technology, EGF-MVSNet network is proposed based on deep learning. First, a feature extraction network incorporating the ECA attention layer is used to improve the network′s attention to the channel feature information; then, an improved combination of GRU modules is used to obtain the GC regularization network for regularization and to reduce the computation of the network; finally, the SmoothL1 loss function and Adam optimizer are used to improve the convergence accuracy at the later stage of model training and to optimize the model′s losses and parameters. Through testing and validation on the DTU public dataset, the EGF-MVSNet network proposed in this paper improves the completeness by 22.1% and the overall model score by 11.5% compared to the MVSNet network, which confirms that the EGF-MVSNet network can significantly improve the quality of the reconstruction results and reduce the network′s consumption of memory. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
