融合注意力与自适应加权特征金字塔的铝型材缺陷检测
Aluminum profile defect detection based on attention and adaptive weighted feature pyramid
赵伟 1刘国华2
作者信息
- 1. 天津工业大学机械工程学院,天津 300387
- 2. 天津工业大学机械工程学院,天津 300387;天津市现代机电装备技术重点实验室,天津 300387
- 折叠
摘要
本文提出了一种新的铝型材缺陷检测方法,以解决现有方法在精度和小目标缺陷检测方面存在的问题.首先在特征提取网络中引入坐标注意力机制(coordinate attention,C A),以防止信息丢失,从而避免缺陷漏检.其次,引入自适应加权特征金字塔(adaptive weighted feature pyramid,AWFPN),优化特征图的感受野和注意力融合,从而提升缺陷特征的捕获和利用效率.同时,改进了边界框回归损失函数,更好地处理缺陷尺度变化和定位误差,从而在提高检测精度的同时加快了模型的收敛速度.实验结果显示,该方法在提升检测效果方面取得显著进展,尤其在脏点和划痕等缺陷的识别能力上有明显提升.
Abstract
This paper proposes a new method for defect detection of aluminum profiles to solve the problems of existing methods in accuracy and small target defect detection.Firstly,the coordinate attention mechanism(CA)is introduced into the feature extraction network to prevent information loss,thereby avoiding missed defect detection.Secondly,the adaptive weighted feature pyramid(AWFPN)is introduced to optimize the receptive field of feature maps and attention fusion,thereby improving the efficiency of defect feature capture and utilization.At the same time,the bounding box regression loss function is improved to better handle the scale change and positioning error of defects,thereby improving the detection accuracy and speeding up the convergence speed of the model.Experimental results show that the method has significantly improved the detection effect,especially in the recognition ability of defects such as dirty points and scratches.
关键词
铝型材/缺陷检测/注意力机制/深度学习/YOLOv5Key words
aluminum profile/defect detection/attention mechanism/deep learning/YOLOv5引用本文复制引用
出版年
2025
NETL
NSTL
万方数据