摘要
现有石英坩埚内壁缺陷检测方式以人工目检为主,准确度低且效率低下。深度学习技术可显著提升工业缺陷检测的精度与效率,同时石英坩埚质检产线终端设备计算资源有限,为此提出了一种轻量化石英坩埚内壁缺陷检测模型QCD-YOLO。在YOLOv8n的基础上利用部分卷积(Partial Convolution,PConv)设计全新的C2f结构,降低模型计算量与参数量;使用ADown降采样模块替换主干网络中Conv模块,提升小目标缺陷检测能力;引入多尺度空洞注意力(Multi-Scale Dilated Attention,MSDA),不增加额外计算成本的情况下高效聚合不同尺度的语义信息;设计Inner-Shape IoU损失函数替换原损失函数。实验结果表明,改进模型在自建石英坩埚内壁缺陷数据集上mAP达到98.1%,相较于原模型YOLOv8n提升1.2%,同时,参数量下降0.83 M,计算量下降2.2 G,权重下降1.58 MB,可满足检测精度要求,同时更容易部署至石英坩埚质检产线。
Current defect detection methods for the inner walls of quartz crucibles primarily rely on manual visual inspection,resulting in low accuracy and efficiency.Deep learning technologies can significantly enhance the precision and efficiency of industrial defect detection.However,the computational resources of the quality inspection terminal equipment for quartz crucibles are limited.Therefore,we propose a lightweight defect detection model,QCD-YOLO,for the inner walls of quartz crucibles.Based on YOLOv8n,we design a novel C2f structure utilizing Partial Convolution(PConv)to reduce the model’s computational and parameter loads.The ADown downsampling module replaces the Conv module in the backbone network to enhance the detection performance for small defects.Additionally,a Multi-Scale Dilated Attention(MSDA)mechanism is introduced to efficiently aggregate semantic information at different scales without increasing computational costs.The Inner-Shape IoU loss function is also designed to replace the original loss function.Experimental results demonstrate that the improved model achieves a mean Average Precision(mAP)of 98.1%on a self-constructed dataset of defects on the inner surface of quartz crucibles,representing a 1.2%improvement over the original YOLOv8n model.Furthermore,the parameter count is reduced by 0.83 M,the computational complexity is reduced by 2.2 G,and the model’s weight size is decreased by 1.58 MB.These improvements meet the accuracy requirements for defect detection while facilitating easier deployment on quartz crucible quality inspection production lines.
作者
赵谦
郭乔峰
尹怡晨
陶涌
黄晶晶
ZHAO Qian;GUO Qiaofeng;YIN Yichen;TAO Yong;HUANG Jingjing(School of Communication and Information Engineering,Xi’an University of Science and Technology,Xi’an 710054,China;Xi’an Dishan Vision Technology Limited Company,Xi’an 712044,China;Chengli New Materials Limited Company,Baotou Inner Mongolia 014060,China)
出处
《激光杂志》
北大核心
2025年第7期69-76,共8页
Laser Journal
基金
陕西省教育厅服务地方企业(No.22JC050)
陕西省科技成果转化计划-百项科技成果转化行动项目(No.2023-YD-CGZH-29)
陕西秦创原“科学家+工程师”队伍建设项目(No.S2023-YF-LLRH-QCYK-0271)
西安市科技计划(科学家+工程师队伍建设项目)(No.24KGDW0030)。
