Scholarlink

<p>Accurate 3D surface reconstruction of glass bottles is crucial for various manufacturing applications, including quality inspection, defect detection and digital twin. Traditional reconstruction methods struggle with transparent and reflective surfaces due to ambiguous depth cues and refraction effects. In this work, a neural implicit representation-based approach to reconstruct the 3D surfaces of glass bottles from RGB images has been proposed, called SGNeu. Firstly, SGNeu constructs the continued SDF using a zero-level set and optimizes surface reconstruction through volume rendering. Secondly, to overcome the interference caused by glass materials, SGNeu incorporates an attention mechanism and designs an auxiliary boundary extraction module to optimize SDF prediction using extracting boundaries of glass bottles. Thirdly, SGNeu proposes to use a separation volume rendering model to separately render glass and non-glass zone. The reconstructed surfaces can be directly utilized for simulation, virtual inspection, and reverse engineering in manufacturing. Experimental evaluations on synthetic and real-world datasets demonstrate that SGNeu achieves superior reconstruction quality compared to existing techniques, particularly in handling transparency and complex lighting conditions. SGNeu provides a practical solution for generating high-fidelity 3D representations of glass bottles, contributing to the advancement of intelligent manufacturing and industrial automation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Surface reconstruction of glass bottles using neural implicit representations for manufacturing system

  • 期刊论文

  • 发表日期：

    2025年8月27日

查看全文
  • Chenyu Tian,
  • Xin Ye,
  • Jiewu Leng,
  • Xinyu Li,
  • Huafeng Ding,
  • Long Wen

摘要

Accurate 3D surface reconstruction of glass bottles is crucial for various manufacturing applications, including quality inspection, defect detection and digital twin. Traditional reconstruction methods struggle with transparent and reflective surfaces due to ambiguous depth cues and refraction effects. In this work, a neural implicit representation-based approach to reconstruct the 3D surfaces of glass bottles from RGB images has been proposed, called SGNeu. Firstly, SGNeu constructs the continued SDF using a zero-level set and optimizes surface reconstruction through volume rendering. Secondly, to overcome the interference caused by glass materials, SGNeu incorporates an attention mechanism and designs an auxiliary boundary extraction module to optimize SDF prediction using extracting boundaries of glass bottles. Thirdly, SGNeu proposes to use a separation volume rendering model to separately render glass and non-glass zone. The reconstructed surfaces can be directly utilized for simulation, virtual inspection, and reverse engineering in manufacturing. Experimental evaluations on synthetic and real-world datasets demonstrate that SGNeu achieves superior reconstruction quality compared to existing techniques, particularly in handling transparency and complex lighting conditions. SGNeu provides a practical solution for generating high-fidelity 3D representations of glass bottles, contributing to the advancement of intelligent manufacturing and industrial automation.