摘要
单阶段目标检测网络特征融合性能不足,且型钢生产现场计算资源受限,导致型钢表面缺陷检测精度较低。针对上述问题,提出一种改进YOLOX-s的轻量级型钢表面缺陷检测算法。首先,提出一种轻量级双路并行注意力模块并将该模块引入YOLOX-s,以提高网络对缺陷特征的敏感度和提升有效特征的提取效率;其次,在Neck中构建双向特征金字塔网络(BiFPN)加权特征融合路径,促进浅层细节特征与深层语义特征的交互融合,强化网络特征融合能力,并在网络中引入深度可分离卷积(DSC)对模型进行轻量化处理;最后,将模型的边界框回归损失函数替换为完全交并比(CIoU)损失,加快模型收敛,提升预测框的定位精度。在NEU-DET数据集上的实验结果表明,所提算法的平均精度均值(mAP)达到了74.6%,比原始YOLOX-s提升了4.8个百分点,推理帧率达到75.2 frame/s,能够满足实时性检测的需求;生产现场采集的型钢数据集进一步验证了所提算法的可行性。
A lightweight steel surface defect detection algorithm based on an improved YOLOX-s network was proposed to address the problem of low accuracy in steel surface defect detection caused by the limited computing resources in steel production sites and insufficient feature fusion performance in single-stage object detection networks.Firstly,a lightweight dual-path parallel attention module was proposed and introduced into YOLOX-s to enhance the sensitivity of the network to defect features and improve the efficiency of effective feature extraction.Secondly,the Bi-directional Feature Pyramid Network(BiFPN)weighted feature fusion path was constructed in the Neck to promote the interaction and fusion of shallow detail features with deep semantic features,strengthening the network’s feature fusion capability,and Depthwise Separable Convolution(DSC)was introduced to lighten the model.Finally,the bounding box regression loss function of the model was replaced to Complete Intersection over Union(CIoU)loss to speed up the convergence of the model and improve the localization accuracy of prediction boxes.Experimental results on the NEU-DET dataset show that the mean Average Precision(mAP)of the proposed algorithm reaches 74.6%,which is 4.8 percentage points higher than that of the original YOLOX-s,and the inference speed reaches 75.2 frame/s,which can meet the real-time detection needs.Finally,the feasibility of the proposed algorithm was further verified using a steel dataset collected from production sites.
作者
黄啸
吴龙
黎尧
吕宏泽
HUANG Xiao;WU long;LI Yao;LYU Hongze(College of Mechanical and Electrical Engineering,Fujian Agriculture and Forestry University,Fuzhou Fujian 350000,China;School of Mechanical and Electrical Engineering,Sanming University,Sanming Fujian 365004,China)
出处
《计算机应用》
CSCD
北大核心
2023年第S02期201-208,共8页
journal of Computer Applications
基金
福建省科技重大专项(2022HZ026025)
福建省2023年揭榜挂帅成果转化项目(2023T0101)
福建省高校创新团队发展计划项目(IRTSTFJ)
中央引导地方科技发展专项(2022L3044,2021L3029)。
关键词
YOLOX-s
双向特征金字塔网络
并行注意力
完全交并比
损失
深度可分离卷积
型钢表面缺陷检测
YOLOX-s
Bi-directional Feature Pyramid Network(BiFPN)
parallel attention
Complete Intersection over Union(DIoU)loss
depthwise separable convolution
steel surface defect detection
