摘要
提出一种金属表面缺陷检测方法的改进模型.首先,基于YOLOv3(you only look once v3)目标检测模型,使用多尺度卷积并行结构,提取、融合多尺度特征;其次,使用高效下采样,在保留特征信息的同时减少特征升维的计算量;最后,使用空间可分离卷积,在保持感受野不变的前提下增加模型的宽度与深度,从而得到模型参数量减少、同时提升了模型性能的改进模型YOLOv3I(you only look once v3 inception).改进模型提高了对复杂缺陷的特征提取能力,并进一步降低了对硬件配置的要求.实验结果表明,改进模型在精度与计算效率上均有明显提升.平均准确率在公开数据集上约提高5%,在企业提供的轴承数据集上约提高3%,模型参数量下降超过20%,两个数据集上模型浮点计算量分别减少1.6×10^(9)和1.2×10^(10)次.
We proposed an improved model of metal surface defect detection method.Firstly,based on the YOLOv3(you only look once v3)object detection model,a multi-scale convolution parallel structure was used to extract and fuse multi-scale features.Secondly,efficient downsampling was used to maintain the feature information and reduce the computation caused by feature dimension raising.Finally,spatial separable convolution was used to increase the width and depth of the model while keeping the receptive field unchanged,so that an improved model YOLOv3I(you only look once v3 inception)with reduced the amount of model parameters and improved the performance of the model was obtained.The improved model improved the feature extraction ability for complex defects and further reduced the requirements for hardware configuration.The experimental results show that the improved model has significantly improved both accuracy and calculation efficiency,with an average accuracy increase of about 5%on the public dataset,and about 3%on the bearing dataset provided by the enterprise.The amount of model parameters decreases by more than 20%,and the floating point computation of the model reduces by 1.6×10^(9) and 1.2×10^(10) times on both two datasets respectively.
作者
刘浩翰
孙铖
贺怀清
惠康华
LIU Haohan;SUN Cheng;HE Huaiqing;HUI Kanghua(College of Computer Science and Technology,Civil Aviation University of China,Tianjin 300300,China)
出处
《吉林大学学报（理学版）》
CAS
北大核心
2023年第3期612-622,共11页
Journal of Jilin University:Science Edition
基金
国家重点研发计划项目(批准号:2020YFB1600101)
天津市教委科研项目(批准号:2020KJ024).
关键词
缺陷检测
特征提取
多尺度卷积并行结构
空间可分离卷积
下采样
defect detection
feature extraction
multi-scale convolution parallel structure
spatial separable convolution
downsampling
作者简介
第一作者:刘浩翰(1966-),男,汉族,硕士,副教授,从事图形图像与可视分析的研究,E-mail:hhliu@cauc.edu.cn;通信作者:孙铖(1994-),男,汉族,硕士研究生,从事图像处理缺陷检测的研究,E-mail:15613125247@163.com.
