Scholarlink

<p>To overcome low accuracy and poor generalization in detecting minute surface defects (Splashes, Scratches, Chipping) on optical wedge plates, this paper proposes a Swin-Unet model that incorporates physical characteristic information of the defects, designated as the PINN-Swin-Unet. The core innovations are twofold: First, leveraging the distinct physical responses of defects under varied spectra, a tri-band light source imaging system is employed to acquire images rich in physical features. These are integrated via a Physics-Informed Feature Extraction (PIFE) module into the skip connection layers of Swin-Unet and propagated to the encoder stage. Second, electromagnetic boundary conditions are incorporated as a physics-informed inductive bias. Instead of a direct analogy to electric field intensity, we model the image gradient as a proxy for the spatial variation of electromagnetic energy density (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(|{\text{E}}|^{2}\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mrow> <mo stretchy="false">|</mo> <mtext>E</mtext> <mo stretchy="false">|</mo> </mrow> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation>) caused by refractive index discontinuities. Based on this, three constraints inspired by Maxwell's equations are introduced: 1. Boundary constraint based on energy density gradient singularities; 2. Internal homogeneity constraint; 3. Background suppression constraint. Guided by these physical constraints, experimental results on an expanded dataset of 716 physical samples demonstrate the superior performance of the PINN-Swin-Unet. It achieves a mean Intersection over Union (mIoU) of 0.873, significantly outperforming the baseline and industrial software. Integrating physical laws as inductive bias enables features with greater generalizability and interpretability, establishing a new paradigm for high-precision optical inspection.</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

Physics-informed deep learning for surface defect detection on optical wedge plates: enhancing accuracy and generalization

  • 期刊论文

  • 发表日期：

    2026年4月7日

查看全文
  • Jishi Zheng,
  • Feifan Lv,
  • Yi Ding,
  • Chenyu Guo,
  • Yan Wang,
  • Dingrong Yi,
  • Pengfei Ma,
  • Xing Sun

摘要

To overcome low accuracy and poor generalization in detecting minute surface defects (Splashes, Scratches, Chipping) on optical wedge plates, this paper proposes a Swin-Unet model that incorporates physical characteristic information of the defects, designated as the PINN-Swin-Unet. The core innovations are twofold: First, leveraging the distinct physical responses of defects under varied spectra, a tri-band light source imaging system is employed to acquire images rich in physical features. These are integrated via a Physics-Informed Feature Extraction (PIFE) module into the skip connection layers of Swin-Unet and propagated to the encoder stage. Second, electromagnetic boundary conditions are incorporated as a physics-informed inductive bias. Instead of a direct analogy to electric field intensity, we model the image gradient as a proxy for the spatial variation of electromagnetic energy density ( \(|{\text{E}}|^{2}\) | E | 2 ) caused by refractive index discontinuities. Based on this, three constraints inspired by Maxwell's equations are introduced: 1. Boundary constraint based on energy density gradient singularities; 2. Internal homogeneity constraint; 3. Background suppression constraint. Guided by these physical constraints, experimental results on an expanded dataset of 716 physical samples demonstrate the superior performance of the PINN-Swin-Unet. It achieves a mean Intersection over Union (mIoU) of 0.873, significantly outperforming the baseline and industrial software. Integrating physical laws as inductive bias enables features with greater generalizability and interpretability, establishing a new paradigm for high-precision optical inspection.