.png){kind=icon}
.png){kind=icon}
Return
Multi-model integrated line detection robust to environmental variations using Bayesian optimization
.png){kind=icon}
OA
AI
DOI：10.1016/j.cie.2026.112343.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
As industrial automation and vision-based inspection technologies become increasingly widespread, the demand for reliable line detection across diverse manufacturing processes continues to grow. However, real industrial environments involve illumination changes, metallic surface reflections, and texture variability, which cause substantial performance fluctuations in single-model line detectors. Moreover, when data collection is limited, learning-based approaches alone may not provide sufficient robustness to such environmental variations. In this study, we propose a multi-model integrated line detection framework that runs multiple detectors operating in independent parameter spaces and combines their complementary outputs using weighted RANSAC. We further apply risk-aware Bayesian optimization with Conditional Value-at-Risk as the objective to automatically derive parameter configurations that are robust to environmental variations. Unlike conventional mean-based optimization, this approach directly targets worst-case performance and, to the best of our knowledge, represents the first application of CVaR-based Bayesian optimization to a computer vision task. The proposed method is validated on 779 real industrial images collected from a shipyard mid-assembly welding environment, achieving up to a 33.5% improvement in CVaR and confirming generalization on a separate held-out validation set.
Keywords:
Line detection
Multi-model integration
Bayesian optimization
Risk-aware optimization
Weighted RANSAC
Industrial inspection
Journal
C
IF:
6.5
Papers:
492
Citations:
0