摘要
针对铸件在铸造过程中易出现漂芯、漏芯等成孔类缺陷,采用视觉检测时铸件与背景特征相近,产生特征网络提取能力欠缺的问题,提出改进YOLOX的铝铸件表面成孔缺陷检测方法。构建成孔缺陷数据集;引入SE注意力机制,提升特征重用率;替换原卷积块为CBH,稳定特征拟合过程;完善CIOU边界框回归损失函数,加速模型收敛。试验证明,改进模型在铝铸件缺陷数据集的平均检测精度提升至97.13%,单图推理速度为0.0162 s,可快速准确地完成铝铸件表面缺陷检测。
Aiming at the problem of floating cores and missing cores that commonly occurs during the casting process,and the lack of feature network extraction capabilities due to the similarity of casting and background features when visual inspection is used,an optimized YOLOX method for detecting pore-forming defects on the surface of aluminum castings was proposed.A hole defect dataset was constructed,and SE attention mechanism was introduced to improve feature reuse rate.Meanwhile,original convolution block was replaced by CBH to stabilize feature fitting process,and CIOU bounding box regression loss function was modified to accelerate model convergence.The results indicate that the average detection accuracy of optimized model in the aluminum casting defect dataset is increased to 97.13%,and the single-image reasoning speed is 0.0162 s,which can quickly and accurately detect aluminum casting surface defects.
作者
胡佳琪
王成军
杨超宇
胡鹏
Hu Jiaqi;Wang Chengjun;Yang Chaoyu;Hu Peng(School of Computer Science and Engineering,Anhui University of Science and Technology;School of Artificial Intelligence,Anhui University of Science and Technology)
出处
《特种铸造及有色合金》
CAS
北大核心
2023年第9期1205-1209,共5页
Special Casting & Nonferrous Alloys
基金
安徽省自然科学基金资助项目(2208085ME128)。
关键词
铝铸件
表面缺陷检测
端到端
深度学习
注意力机制
Aluminum Casting
Surface Defect Detection
End-to-End
Deep Learning
Attention Mechanism
