摘要
Abstract
Organic board is a high-strength,high-stiffness,lightweight composite material based on the pultrusion molding process,which is widely used in automotive manufacturing materials because of its excellent mechanical properties.Appearance defects such as color abnormality,scratches,stains,pits,parting seams,overflow,etc.,are prone to occur in the production process,resulting in unqualified quality control leading to customer complaints and rework,and increasing production costs.The existing production process mainly relies on manual visual inspection,which is a strong subjectivity and inefficiency.Therefore,a new method that replaces manual real-time detection is needed.A YOLOv5 defect target detection model YOLOv5_CBAM,based on the attention mechanism,has been designed to achieve detection of appearance defects in organic boards.Based on the characteristics of continuous production of organic boards using Dalsa line array camera real-time acquisition of organic board image data in production,input to the YOLOv5 model to extract the defective features in the image,the backbone network introduces the attention mechanism CBAM to enhance the model of the defects of the key features of the attention of the channel and spatial weighting to improve the accuracy and robustness of the model.Experiments show that the mAP of the improved model YOLOv5_CBAM on the validation set of organic plate dataset is 98.6%,which is 4.9 percentage points higher than the original YOLOv5s,and the mAP is 23.4,12.4 and 11.5 percentage points higher compared with the Faster RCNN,YOLOv3 and YOLOv4 models,respectively,and the model's single-image inference time is 41 ms.The experimental results show that the improved model YOLOv5_CBAM can accurately detect the appearance defects of organic boards in real time.关键词
缺陷检测/复合材料/YOLO/机器视觉/有机板Key words
defect detection/composites/YOLO/machine vision/organic plates分类
信息技术与安全科学
