Analysis on main influence factors for silicon rod defects detection model using machine learning

The defects detection for silicon rod surface is accessible to data volume, detection model, light conditions and other factors, which affects the accuracy of results. In this paper, a flaw detection test system for silicon surface is built based on the framework of YOLOv5. An actual industrial test and experimental analysis are conducted to study the influence of data volume, light conditions, detection model on the performance of detection system. It is found that the data can be reduced by 27M bytes using YOLOv5s detection model compared with using of YOLOv5m model. The raw defects data can be increased to 3.12 times the original level by data enhancement. The optimal length of strip light source is 800-1000mm, and the lighting uniformity is 0.0231–0.0256. The experimental results show that the inference time for a single image is reduced by 30ms and the recall rate is improved by 0.01 when under YOLOv5s detection model. The defect detection rate is improved by 6.9%, with multi-check rate reduced by 3.2% and missed detection rate redcued by 0.7%, wheh the confidence level is 0.4. The experimental results of influence factors on the defects detection model provide a foundation for the detection during the production of silicon rod.