Abstract: To address the limitations of existing studies on silicon rod surface defect detection,as well as the issues of misdetection and imbalanced detection performance among different defect types under a unified confidence threshold, a multi-confidence-based surface defect detection system for silicon rods is proposed. Firstly,the mechanical structure of the system is designed with a diffuse reflective structured light illumination strategy. By integrating a strip light source with a white fiber-optic curtain,the system effectively resolves illumination challenges caused by the strong reflectivity and large dimensions of silicon rods. Secondly,a precise evaluation metric framework is constructed,centered on Accurate Detection Rate (ADR),Missed Detection Rate (MDR),and Over Detection Rate (ODR). Based on multi-view image acquisition of single rods and a dual-matrix discrimination method,the system establishes a regulation framework linking defect characteristics,evaluation metrics,and optimal confidence thresholds,enabling the identification of detection accuracy and false detection patterns. The proposed dual-matrix approach enables a comprehensive quantitative assessment of the system′s effectiveness across all perspectives. Finally,a defect-type-specific confidence segmentation method is introduced.According to variations in imaging features among defect types,optimal confidence thresholds are assigned to 8 types of defects,including pits and edge breakage,thereby constructing a highly efficient multi-confidence detection model. Experimental results demonstrate that the proposed model significantly improves the ADR across various defect types. In particular,the ADR for pits,bright spots,watermark,and smudges increased respectively by 8.1 %,4.8 %,2.2 % and 3.5 %. Additionally,most defects' exhibit decreased ODR and MDR,with notable reductions in the ODR for pits,bright spots,watermark,and smudges,and in the MDR for crystal cracks and bright spots. The results confirm that the multi-confidence detection system′s ability to accurately locate and identify defects,effectively overcoming prior research limitations and providing a robust and efficient solution for silicon rod quality control.

Key words: silicon rods, surface defect detection, multiple confidence, fill light method