基于卷积神经网络的表面划痕识别方法研究  

Research on Recognition Method of Scratch on Surface Based on Convolution Neural Network

作　　者：季昌灿 杨立拥 顾磊 刁亦冰 张宇[1] 赵子健 王奇瑞 JI Changcan;YANG Liyong;GU Lei;DIAO Yibing;ZHANG Yu;ZHAO Zijian;WANG Qirui(School of Aeronautical and Mechanical Engineering,Changzhou Institute of Technology,Changzhou 213032,China;Hai′an Institute of Intelligent Equipment,SJTU,Nantong 226600,China)

机构地区：[1]常州工学院航空与机械工程学院,江苏常州213032 [2]海安上海交通大学智能装备研究院,江苏南通226600

出　　处：《河南科技》2024年第22期21-26,共6页Henan Science and Technology

基　　金：重点领域科技攻关计划(2023AB015)。

摘　　要：[目的]传统的人工目测方法已无法满足对金属表面划痕进行快速、准确和自动化检测的需求。因此,基于数据集的构建和卷积神经网络模型的搭建,提出了复杂形状表面划痕识别的方法。[方法]首先,创建金属表面划痕的数据集。其次,设计并训练一个基于Yolov8n的卷积神经网络模型。该模型包括主干网络、头部网络和颈部网络,可应对不同划痕的识别需求。[结果]在模型训练完成后,F1曲线在0.3~0.5达到最优,表明该模型在处理各种划痕时具有良好的泛化能力。通过PR曲线分析,当精确率为0.65、召回率为0.8时,该模型的预测效果最佳。[结论]模型优化为金属表面划痕的自动检测和识别提供了有效的技术支持,具有实际应用价值。[Purposes]In view of the fact that the traditional manual visual inspection method cannot meet the needs of fast,accurate and automatic scratch detection on metal surfaces.Therefore,based on the con-struction of data sets and the convolutional neural network models,a method for scratch recognition of complex shape surfaces is proposed.[Methods]First,a data set of metal surface scratches was created;then,a convolutional neural network model based on Yolov8n is designed and trained.The model in-cludes backbone network,head network and neck network,which can meet the recognition requirements of different scratches.[Findings]After the training of the model,the F 1 curve was optimal in the inter-val of 0.3~0.5,which indicated that the model had good generalization ability in dealing with various scratches.Through the analysis of PR curve,when the accuracy rate is 0.65,and the recall rate is 0.8,the prediction effect of the model is the best.[Conclusions]The model optimization provides effective technical support for the automatic detection and recognition of metal surface scratches,and has practical application value.

关 键 词：金属表面划痕 目标检测 YOLO算法 数据集 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]