基于轻量级YOLOv8的坩埚缺陷识别算法  

Crucible defect recognition based on lightweight YOLOv8

作　　者：罗彬彬 庞晴蔚 康希平 LUO Binbin;PANG Qingwei;KANG Xiping(Changsha Kaiyuan Instrument Corporation Ltd.,Changsha 410100,China)

机构地区：[1]长沙开元仪器有限公司,湖南长沙410100

出　　处：《煤质技术》2025年第2期85-90,96,共7页Coal Quality Technology

基　　金：岳麓山工业创新中心衡山实验室科研基金资助项目(HST2312)。

摘　　要：基于轻量级YOLOv8的坩埚缺陷识别方法,可解决机器人智能化验系统中坩埚几何形态多样性显著、缺陷样本分布失衡导致的检测可靠性难题。首先基于DDPM的图像数据扩充增强策略,根据数据标签分布,实现坩埚图像数据的平衡扩充;随后为了更好地平衡网络模型的推理速度和检测精度,提出改进的YOLOv8坩埚缺陷识别算法,引入无参数注意力和自适应加权机制。实验结果表明,该方法的平均精度(mAP)为97.40%,单张图像的推理时间为1.20 ms。尽管推理速度比YOLOv11-n延迟0.2 ms,但当前通用目标检测方法中的轻量级方法检测精度居于最优,所提方法能有效应用于工业坩埚缺陷检测场景。To address the challenges of low detection reliability in robotic intelligent laboratory systems caused by significant geometric diversity of crucibles and imbalanced defect sample distribution,this paper proposes a crucible defect detection method based on the lightweight YOLOv8.Firstly,an image data augmentation strategy using Denoising Diffusion Probabilistic Model(DDPM)is introduced to balance and expand crucible image data according to the distribution of data labels.Subsequently,to better balance the inference speed and detection accuracy of the network model,an improved YOLOv8 crucible defect detection algorithm is proposed,incorporating a non-parametric attention mechanism and an adaptive weighting scheme.Experimental results demonstrate that the proposed method achieves a mean Average Precision(mAP)of 97.40%with an inference time of 1.20 ms per image.Although the inference speed is 0.2 ms slower than YOLOv11-n,the detection accuracy is optimal among current general object detection methods,especially lightweight ones.The proposed method can be effectively applied in industrial crucible defect detection scenarios.

关 键 词：坩埚缺陷识别算法 机器人智能化验系统 YOLOv8 数据扩充 目标检测 加权机制 推理时间 

分 类 号：TQ533.9[化学工程—煤化学工程]