基于邻域表面特征的隐式神经表示方法
Implicit neural representation method based on neighborhood surface features
于楚飞 1苏工兵 1王晶 1袁梦 1曾文豪1
作者信息
- 1. 武汉纺织大学机械工程与自动化学院,湖北武汉 430200;武汉纺织大学湖北省数字化纺织装备重点实验室,湖北武汉 430200
- 折叠
摘要
隐式神经表示能够实现任意分辨率下的表面重建.现有方法仅使用了坐标信息,未考虑到其邻域表面上的特征对局部形状的贡献,因此难以精确恢复表面上复杂的纹理和拓扑.为此,提出一种采用编码-解码策略的改进模型,提高表面重建的精度.编码器获取坐标在邻域表面上的特征编码和其在高维空间下的映射编码.在解码器中应用损失自适应加权策略,提高编码信息的利用率.实验结果表明,较现有方法显著提高了重建结果的精度,其重叠度和F-score分别提高了1.458％和1.46％,平均倒角距离降低了 0.08.
Abstract
Implicit neural representations can achieve surface reconstruction at arbitrary resolutions.Existing related methods on-ly use coordinate information,without considering the information contained in the features on the neighboring surface,making it difficult to accurately reconstruct the complex texture and topology of the surface.To address this,an improved model using an encoder-decoder approach was proposed to improve the accuracy of surface reconstruction.The encoder obtained feature encoding of the coordinates on the neighboring surface and their mapping encoding in high-dimensional space.The decoder applied a loss-adaptive weighting strategy to improve the utilization of the encoding information.Experimental results show that using this method,the accuracy of the reconstruction results is significantly improved,the intersection over union and F-score are increased by 1.458％and 1.46％respectively,and the average chamfer distance is decreased by 0.08,compared to existing methods.
关键词
表面重建/隐式神经表示/符号距离函数/点云/多层感知机/傅里叶变换/编码器Key words
surface reconstruction/implicit neural representation/signal distance function/point cloud/multilayer perceptron/Fourier transform/encoder引用本文复制引用
基金项目
湖北省科技专项基金项目(2019AEE011)
湖北省数字化纺织装备重点实验室开放基金项目(KDTL2022011)
出版年
2024
NETL
NSTL
万方数据