Profile extraction and defect detection for stereolithography curing process based on multi-regularized tensor decomposition

Yinwei Zhang, Tao Zhang, Jian Liu, Wenjun Kang, Rongguang Liang, Barrett G. Potter

Research output: Contribution to journalArticlepeer-review


Optical lenses cured from the stereolithography process are still at their primitive stage, where the detection of process faults and product defects is of great importance. Such needed capability is enabled by an in situ process monitoring system with advanced camera sensors that collect high-dimensional images of the outer geometric profile and/or inner defects of the cured lenses. The state-of-the-art methods fall short in distinguishing the profile of the cured lens and the sparsely distributed spots from the curable resin shown in noisy images, leaving profile assessment and defect detection ineffective and unreliable. In this paper, a Multi-Regularized Tensor Decomposition (MRTD) Method is proposed to improve the performance of profile extraction and defect detection. It decomposes an in situ image of the curing process into four components: the smooth background, the smooth cured lens profile, the sparse spots, and the white noises. A modified ADMM algorithm is developed to solve this decomposition as a large-scale optimization problem. A real-world case study is conducted to demonstrate that the proposed method outperforms existing state-of-the-art methods with respect to profile extraction accuracy and defect detection sensitivity.

Original languageEnglish (US)
Pages (from-to)100-111
Number of pages12
JournalJournal of Manufacturing Systems
StatePublished - Jun 2024


  • ADMM algorithm
  • Additive manufacturing
  • Image processing
  • Optimization
  • Tensor decomposition

ASJC Scopus subject areas

  • Control and Systems Engineering
  • Software
  • Hardware and Architecture
  • Industrial and Manufacturing Engineering


Dive into the research topics of 'Profile extraction and defect detection for stereolithography curing process based on multi-regularized tensor decomposition'. Together they form a unique fingerprint.

Cite this