YoloGT:基于ViT的钢板表面轻量级瑕疵检测算法  

YOLOGT:METAL SURFACE LIGHTWEIGHT DEFECT DETECTION ALGORITHM BASED ON VIT

作　　者：李宁 刘青 熊俊 尚英强 时晨杰 Li Ning;Liu Qing;Xiong Jun;Shang Yingqiang;Shi Chenjie(Beijing Zhuoyue Power Construction Co.,Ltd.,Beijing 100027,China;State Grid Beijing Cable Company,Beijing 100073,China)

机构地区：[1]北京卓越电力建设有限公司,北京100027 [2]国网北京电缆公司,北京100073

出　　处：《计算机应用与软件》2025年第3期319-325,340,共8页Computer Applications and Software

摘　　要：目前基于传统方法设计的瑕疵检测算法速度较慢、检测精度低,而基于深度学习的瑕疵检测技术,也因其庞大的计算量和模型体积,导致无法在有限的设备上进行部署使用。针对以上问题,引入Vision Transformer(ViT)结构至特征提取部分,并利用一种改变卷积结构的算法(GhostNet),实现瑕疵检测网络结构的轻量化,得到网络模型Yolov5s-Ghost-ViT(YoloGT)。与原模型相比,YoloGT模型体积、计算量和参数量分别减小了42.4%、47.9%、38.8%,mAP值在VOC和NEU数据集上分别提升1.65百分点和2.9百分点。相比原算法,更适用于工业场景中钢板表面瑕疵的嵌入式实时检测系统。At present,defect detection algorithms based on traditional methods are slow and have low detection accuracy.However,the defect detection technology based on deep learning is also unable to be deployed and used on limited devices due to its huge calculation volume and model size.In response of the above problems,the Vision Transformer(ViT)structure was introduced to the feature extraction part and an algorithm that changed the convolution structure(GhostNet)was used to reduce the weight of the defect detection network structure,and the network model Yolov5s-Ghost-ViT(YoloGT)was obtained.Compared with the original model,the YoloGT model volume,calculation amount and parameter amount were reduced by 42.4%,47.9%,and 38.8%,respectively,and the mAP value has increased by 1.65 percentage points and 2.9 percentage points on the VOC and NEU data sets,respectively.Compared with the original algorithm,the proposed algorithm is more suitable for the embedded real-time detection system of steel plate surface defects in industrial scenes.

关 键 词：瑕疵检测 YOLOv5 机器视觉 深度学习 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]